A system and a computer implemented method for an agronomical environment

Abstract

The disclosure relates to providing agronomical data for re-training a pre-trained transformer-based model for an agronomical task. The disclosure further relates to using the trained transformer-based model for controlling and / or monitoring an agronomical environment.

Description

[0001] A SYSTEM AND A COMPUTER IMPLEMENTED METHOD FOR AN AGRONOMICAL ENVIRONMENT

[0002] TECHNICAL FIELD

[0003] The claimed invention relates to the technical field of agronomy. In particular, the claimed invention relates to a computer implemented method for releasing a trained transformer-based model for controlling and / or monitoring an agronomical resource, a computer implemented method for using a trained transformer-based model for controlling and / or monitoring an agronomical resource, a computer program product comprising computer readable instructions that when executed on a computer cause the computer to execute one of the methods, a computer-readable storage medium storing computer-readable instructions that when executed on a computer cause the computer to execute one of the methods.

[0004] BACKGROUND

[0005] A transformer-based model may be trained on big data (e.g., unspecific text and image data). Trained transformer-based models have improved capabilities for predicting data patterns such as patterns in natural languages. The improved capabilities are typically attributed to large number of parameters obtained by said training. For example, transformerbased models such as OpenAI GPT may include 117 million parameters (GPT-1), 1.5 billion parameters, 175 billion parameters (GPT-3), 170 trillion parameters (GPT-4). Said parameters allow said GPT models generating improved data output as compared to other models such as recurrent neural networks (RNNs) or long short-term memory (LSTM) networks that do not comprise a transformer component.

[0006] RNN and LSTM were used for analyzing text data before the emergence of transformer architectures. Transformer-based large language models have been found to outperform RNNs and LSTMs for natural language processing, NLP.

[0007] Particularly, transformer-based architectures have advantages over RNNs and LSTMs due to long-range dependencies and parallelization of computation. Furthermore, transformer-based architectures may be pre-trained on larger text-based datasets and fine-tuned (re-trained) for specific tasks with smaller (labeled) datasets. This transfer learning approach has led to state-of-the-art results on numerous NLP tasks, surpassing RNNs and LSTMs.

[0008] Said fine-tunning or re-training may allow using the outstanding analytic capabilities of pre-trained transformer-based models for analyzing data patterns other than patterns in natural languages.

[0009] Thus, one problem associated with the prior art may be formulated as how to use pre-trained transformer-based models for other purposes than natural language processing. Another related problem associated with the prior art is how to obtain data suitable for re-training a pre-trained transformer-based model for a technical purpose. The amount of training data necessary to fine-tune a pre-trained transformer-based model may depend on the specific task, the complexity of the model, and the desired performance level. In many cases, transformer models can be finetuned (re-trained) with smaller amounts of purpose-specific training data as compared to the size of pre-training dataset. However, if the specific application is very different from the pretraining data, re-training a pre-trained transformer-based model may require larger datasets as compared to the scenario when the specific application is similar to the pre-training data. For example, text classification or sentiment analysis may require few hundred MB to a few GB of labeled data. Machine translation may require tens to hundreds of GB of text data for re-training. Question-answering may require several GBs of training data for re-training a pre-trained transformer-based model.

[0010] With better quality of training data, the amount of training data may be reduced (the definition of the term “data quality” is, for example, given in ISO / IEC 20546:2019(en), ISO 8000-66:2021 (en) / Data quality, ISO / IEC DIS 5259-1 (en)). Preprocessing of data for generating high quality training data may involve labeling data, removing noise, irrelevant information and / or alike. Thus, pre-processing data for generating high quality training data may be beneficial for more efficient re-training.

[0011] Larger amounts of training data may allow fine tuning or re-training pre-trained transformer-based models with more parameters. For example, in order to fine-tune Distill BERT model less training data may be required as compared to the amount of data necessary for fine-tunning GPT-4 model. Larger amount of said parameters may provide improved capabilities in data analysis (e.g., GPT-4 is more powerful in analyzing data patterns than GPT-3).

[0012] Pre-trained transformer based models are known in the prior art, for example, ChatGPT (GPT-2, GPT-3, GPT-3-5-turbo, GPT-4), Davinci, BERT (Bidirectional Encoder Representations from Transformers), DistilBERT, Transformer-XL, XLNet (extreme Language understanding Network), T5 (Text-to-Text Transfer Transformer), RoBERTa (Robustly Optimized BERT approach), ELECTRA (Efficiently Learning an Encoder that Classifies Token Replacements Accurately), Reformer, Longformer, DeBERTa (Decoding-enhanced BERT with disentangled attention). The properties, and thus, output data of said transformer-based models may vary for the same input data due to differences in architectures and / or pre-training datasets. Thus, one or more of said transformer-based models (GPT-2, GPT-3, GPT-3-5-turbo, GPT-4, Davinci, BERT, DistilBERT, Transformer-XL, XLNet, T5, RoBERTa, ELECTRA, Reformer, Longformer, DeBERTa) may be used as one alternative of a pre-trained transformer-based model for a technical purpose or one or more of said models may be used in a combination for a technical purpose to provide a plurality of data outputs for complimentary data analysis.

[0013] Alongside with unsurpassed advantages, there are also risks known in the prior art associated with using pre-trained transformer-based models due to that said models may generate spurious or even harmful results. PROBLEM

[0014] It may be an object to improve control and monitoring of an agronomical resource.

[0015] SOLUTION

[0016] The solution to the above problem includes providing agronomical data for re-training a pre-trained transformerbased model for agronomical resources.

[0017] The solution further comprises re-training a pre-trained transformer-based model for providing and / or releasing a trained transformer-based model suitable for agronomical resources. The trained transformer-based model may then be used for controlling and / or monitoring an agronomical resource.

[0018] It may be possible to receive a pre-trained transformer-based model and add re-trained data to the pre-trained model. Adding may be executed in such a way to keep the pre-trained and fine-tuned data separately. In such a solution it may be possible to update the re-trained data and the pre-trained data separately and when a new version of a pre-trained model may arrive, the pre-trained model may be replaced by the new version of the pretrained model substantially without losing the re-trained component. Re-training may comprise replacing one layer of a pre-trained model.

[0019] In an example, the training may comprise adding an additional data layer to the pre-trained model or the prediction device, wherein the additional data layer may comprise weights which may be adjusted during the training process. The weights may be used in addition to weights of the pre-trained which may be left unchanged during the training process. In another example weights of the pre-trained may be adjusted during the training process and no additional data layer may be added. The weights may be adapted to the repository description data and allow for indicating data retrievable from the repository by using the encoded user identifier. In addition, or as an alternative, the training may comprise training for specific agricultural knowledge and / or an agronomical resource. The agricultural knowledge may be provided as structured or unstructured data. The training data may also comprise a list of repositories and / or tools that may be used by the prediction device. The training data may also comprise examples of situations for which specific tools and / or repositories may be used and how they may be used. This may comprise a syntax of how to formulate a query, e.g. of how to transfer a mathematical problem into a mathematical formula.

[0020] The solution also comprises providing a transformer-based model and pre-training said model on generic data such as text-based data, API data, interface data, interface description data, description data of at least one tool and / or of at least one repository, any type of data that may be converted to text data, structured data and / or unstructured data. The pre-training data may also comprise publicly available data associated with agronomical resources and / or agronomic knowledge, such as, for example, data available on websites of agronomical facilities.

[0021] Using any other large language model (LLM), e.g. a foundation model such as a transformer-based model, that may be pre-trained on a big dataset e.g., unspecific, publicly available dataset, videos and / or transcripts of videos and trained, i.e. fine-tunned and / or re-trained, on a smaller dataset and / or more specific data set, e.g., specific dataset, e.g., master data, API data, description data and / or, data structure of at least one tool and / or description data of at least one repository, in order for the trained model to be suitable for controlling and / or monitoring an agronomical source is also withing the scope of the claimed invention.

[0022] The term “monitoring” may also comprise providing an answer to a query to a respective resource.

[0023] ADVANTAGEOUS TECHNICAL EFFECTS

[0024] A pre-trained transformer-based model pretrained on larger datasets, e.g. on publicly available data, unspecific data, and re-trained on smaller datasets i.e., agronomical data or agronomical datasets, is applicable for analyzing agronomical data patterns, identifying solutions for the data patterns, and based on the analysis of the data patterns, generating operating instructions for improving controlling and / or monitoring of agronomical resources.

[0025] In an example the agronomical resource is selected from the group of agronomical resources consisting of a farm, a field, a plant treatment product, a product registration, and an application device.

[0026] Providing agronomical training data enables re-training a pre-trained transformer-based model for agronomical resources. In particular private agronomical know-how may be trained into a re-trained transformer-based model, in order to benefit the general knowledge of the pre-trained model but also of the very specific knowledge, e.g. from private documents administrated for example in master data (MD).

[0027] Pre-processing raw agronomical data, e.g., sorting, filtering, labeling, structuring data, for generating agronomical training data improves the quality of training data, and thus, the training efficiency in that less training data and computing power may be required for re-training or fine-tunning a pre-trained transformer-based model.

[0028] A one or more alternative pre-trained transformer-based model may be used as long as said one or more models are retrainable for an agronomical resource. Said one or more alternative or additional models may have different properties based on different set of parameters obtained during pre-training. Thus, said one or more additional or alternative models may be of interest as a basic model for agronomical inquiries for analyzing data patterns in input data relating to the agronomical resource.

[0029] Trigger based re-training of a trained transformer-based model, scheduled or continuous re-training may result in a high performance of the model in that the model may be trained on the most up to date data. A feedback score may be provided as a quality indicator of generated output data.

[0030] Using the trained transformer-based model for controlling and / or monitoring an agronomical resource such as a field and / or an application device may result in a high efficiency of agronomical resources since many aspects causing inefficient agronomic resources may be identified by the model by analyzing data patters in input data based on agronomical data. Based on the analysis, a solution of how to improve the agronomical resource may be generated by the model in a format of operating instructions, e.g. machine-readable instructions, for controlling and / or monitoring the agronomical resource. A result may also be a feedback how to improve the performance of an agronomical resource, such as how to improve the yield of a field. Machine readable instruction may also be a JSON (JavaScript Object Notation) file that can be converted in a text output, a graphical output and / or a speech / audio output. Reviewing said operating instructions may provide safe integration of the model in an agronomical environment.

[0031] SUMMARY

[0032] The claimed invention is defined by the features of independent claims. Additional embodiments are defined by the features of the dependent claims. The problems inherent to the prior art are solved by the subject matter of the independent claims. Additional synergetic technical effects are provided by the features of dependent claims.

[0033] I n an aspect the disclosure relates to a computer implemented method for releasing a trained transformer-based model for controlling and / or monitoring an agronomical resource, the method comprising: providing, via a computer interface, agronomical training data associated with one or more agronomical tasks, e.g. a task related to the agronomical resource; providing, via the computer interface, a pre-trained transformer-based model comprising at least a transformer component; prompting the pre-trained transformer-based model to re-train the transformer-based model using the agronomical training data associated with the agronomical resource; releasing the trained, in particular the re-trained, transformer-based model for controlling and / or monitoring the agronomical resource. After re-training the pre-trained model may be fine-tuned to specific technical fields, e.g. to the technical field of agronomy. The trained transformer-based model may be adapted to handle an agronomical task and / or provide specific machine-readable instructions and / or output data that may be focused on specific agronomical questions, queries and / or tasks provided by a user. After re-training the pre-trained model may be more focused to a specific agronomical task than before re-training. The re-training may comprise providing description data of at least one tool and / or of at least one repository usable by the re-trained transformer based model in order to generate machine-readable instructions.

[0034] In another aspect the disclosure relates to a computer implemented method for using a trained transformer-based model for controlling and / or monitoring an agronomical resource, the method comprising: releasing a trained transformer-based model for controlling and / or monitoring an agronomical resource; receiving, by a computer processor, access to the trained transformer based model; receiving, via a computer interface, input agronomical data associated with the one or more agronomical tasks; prompting the trained transformer-based model to analyze the input agronomical data and provide operating instructions for controlling and / or monitoring the agronomical resource.

[0035] In an aspect of the disclosure a computer program product is described comprising computer readable instructions that when executed on a computer cause the computer to execute one of the methods.

[0036] In an example, the trained transformer-based model may be provided and / or released after re-training a pretrained transformer based model with agronomical training data. The re-trained transformer-based model may be able to react to specific agronomical questions that may be ignored and / or answered differently by the pre-trained transformer model, without re-training.

[0037] In an example, the pre-trained transformer-based model may not know about description data of at least one tool and / or of at least one repository.

[0038] In a further aspect of the disclosure a computer-readable storage medium may be described storing computer- readable instructions that when executed on a computer cause the computer to execute one of the disclosed methods.

[0039] In yet another aspect of the present disclosure a computer product is described comprising a computer readable token for accessing agronomical training data and / or the trained or the pre-trained model of any of the disclosed methods. In a further aspect of the disclosure, a cloud computing apparatus may be provided comprising a processor for loading computer readable instructions that when executed on the processor cause the processor to execute one of the disclosed methods.

[0040] The cloud computing apparatus, e.g. a server, may be a private network environment which prevents an access from an external source which may not have an authorization and / or an access token.

[0041] DESCRIPTION OF THE DRAWINGS

[0042] In the following, the present disclosure is further described with reference to the enclosed figures. The same reference numbers in the drawings and this disclosure are intended to refer to the same or like elements, components, and / or parts.

[0043] FIG. 1 illustrates an agronomical environment.

[0044] FIG. 2A illustrates an operating system of the agronomical environment of FIG. 1 .

[0045] FIG. 2B illustrates training a pre-trained transformer-based model on agronomical data generated by the agronomical environment of FIG. 1 .

[0046] FIG. 2C illustrates an embodiment of using a trained transformer-based model for controlling and / or monitoring the agronomical environment shown in FIG. 1.

[0047] FIG. 3 illustrates a pre-processing engine of the operating system shown in FIG. 2.

[0048] FIG. 4 illustrates an agronomical data structure generated by the agronomical environment of FIG. 1 .

[0049] FIG. 5 illustrates data structure from a user.

[0050] FIG. 6 illustrates further details of re-training a pre-trained transformer-based model for agronomical resources in addition to the embodiments in FIGs. 2a and 2b.

[0051] FIG. 7 illustrates contextualization of prompts. FIG. 8 illustrates trigger-based re-training of a trained transformer-based model when the model is in use for controlling and / or monitoring the agronomical resource.

[0052] FIG. 9 illustrates an embodiment of training an embedding layer.

[0053] FIG. 10A illustrates an embodiment of a transformer encoder architecture.

[0054] FIG. 10B illustrates an embodiment of a transformer decoder architecture.

[0055] FIG. 10C illustrates an embodiment of a transformer encoder-decoder architecture.

[0056] FIG. 11 illustrates an embodiment of training and / or deploying the transformer encoder, the transformer decoder and / or the transformer encoder-decoder.

[0057] FIG. 12 illustrates an embodiment of input embedding.

[0058] DETAILED DESCRIPTION

[0059] An “engine” in the context of FIGs. 1-8 comprises at least one computer processor.

[0060] A “computer interface” in the context of the current disclosure may be, e.g., a graphical user interface, an application programming interface, a web-based interface.

[0061] Terms “training”, “re-training” and “fine tuning” may be used interchangeably in the context of the current application when a pre-trained transformer-based model pre-trained for a first purpose (e.g., analyzing unspecific text-based data) is trained / re-trained / fine-tuned for a second purpose (e.g., an agronomic task). A trained transformer-based model suitable for the second purpose may be further re-trained / fine-tunned for the second purpose in order to improve output data provided by the model.

[0062] "Agronomical environment" or “agronomical resource” may refer, without limitation, to any agronomical infrastructure that is used for an agricultural purpose of generating of one or more agronomical products. The agronomical resource is selected from the group of agronomical resources consisting of a farm, a field, a plant treatment product, a crop protection product, a product registration, and an application device. The agronomical resource may also comprise description data of at least one tool and / or of at least one repository comprising information about the agronomical resource. The agronomical resource may also comprise field performance maps and / or field yield maps, for example as geospatial information.

[0063] The “agronomical product” produced by the agronomical environment may, for example, be any yield, such as yield of a crop.

[0064] The agronomical environment may comprise equipment such as any one or more of an application device, spraying device and / or drone. It may also comprise actuators related to the agronomical resource. In an example a central agronomical management device may be provided, e.g. a field management device, which may collect agronomical data related to a user and / or an agronomical environment of the user.

[0065] Further, the agronomical environment may comprise a plurality of sensors and at least one control system for controlling at least one parameter related to the agronomical resource. This parameter may include a dosage for a plant treatment product like an herbicide, fungicide, or pesticide. Such control functions are usually performed by the control system or controller in response to at least one measurement signal from at least one of the sensors. The control may also be a result of an inquiry of a repository. The plurality of sensors may be distributed in the agronomical environment for monitoring and / or controlling purposes. Such sensors may generate a large amount of data. The sensors may or may not be considered a part of the agronomical resource. An agronomical environment may produce a large amount of agronomical data. In addition to data from sensors, also data about agronomical knowledge may be provided which may indicate different measures to increase and / or decrease the performance of the agronomical resource. The collected data may be stored in a single repository and / or in a plurality of repositories. The data may be classified as personal data (PI I) or publicly available data. Pll data may be private data and may only be accessible after an authorization. Description data of respective repositories may indicate the access method to the respective data. The sensor data may be user specific data and may be assessed as private data allocated to a specific user.

[0066] Said sensors may be used for measuring one or more agronomical parameters and / or for measuring agronomical conditions of said equipment or parameters related to the equipment.

[0067] The repositories may also comprise the description data of at least one tool and / or of at least one repository comprising information about the agronomical resource. In other words, agronomical data and / or agronomical training data comprise description data of at least one tool and / or of at least one repository comprising information about the agronomical resource. For example, the sensors may be used for measuring an agronomical parameter such as a temperature, a level inside a tank, a temperature of a soil. Some examples of sensors based on the parameter that they sense may comprise: temperature sensors, pressure sensors, radiation sensors such as light sensors, LAI (Leaf Area Index) sensors, GAI (Green Area Index) sensors, Infrared sensors, Lidar (Light detection and ranging) sensors and / or a camera.

[0068] The agronomical environment may be a plurality of agronomical environments. I n an example a user or a farmer may control a plurality of fields.

[0069] Besides using sensor data, it may also be possible to receive special knowledge in form of master data (MD) and user specific data associated with the agronomic resource allocated to a specific user. Furthermore, description data for available tools and repositories may be provided. The description data comprise a repository description selected from the group of repository description data consisting of an API description, a GQL (Graphical Query Language) description, and a CFD (Cross Farm Dashboard) description.

[0070] In an embodiment a repository having user specific data associated with the agronomic resource may comprise structured data, e.g. Elasticsearch, RDBMS (Relational Database Management System) and / or unstructured data like documents, e.g. PDF documents.

[0071] In a further embodiment the repository having user specific data associated with the agronomic resource may comprise sensor data, geospatial information, map information and / or weather data.

[0072] The terms “agronomical data” or “agronomical-based data” may be used interchangeably and may relate to data of the resource. Agronomical data may refer to data comprising values, for example, numerical or binary signal values, measured during an agronomical process and / or measurement results of experiments. The results of agronomical experiments may be collected in master data. Master data (MD) may also comprise information about a strategy how to handle a field in a particular situation and / or state like an underperforming state. The agronomical data may be time-series data of one or more of the agronomical parameters and / or the equipment operating conditions. Typically, the agronomical-based data may comprise temporal information of the agronomical parameters and / or the equipment operating conditions, e.g., the data contain time stamps for at least some of the data points related to the agronomical parameters and / or the equipment operating conditions. The agronomical data may comprise time-space data, i.e., temporal data and the location or data related to one or more equipment zones that are located physically apart, such that time-space relationship can be derived from the data. The temporal information may be stored in an event-based database and / or in an event-based repository.

[0073] "Agronomical parameters" may refer to any of the variables related to an agronomical task, for example any one or more of temperature, pressure, time, level, etc. relevant for the agronomical task and / or for querying information. The above definitions of an agronomical environment, agronomical task and the data generated by the agronomical task are mere examples and should not be construed as limiting. It may be understood that the system and the method of releasing a trained transformer-based model for controlling and / or monitoring an agronomical resource may apply to any kind of agronomical task. Any type of agronomical data may be pre- processed by a pre-processing unit to generate the required agronomical training data, e.g. labeled data, prestructured data, structured data, filtered data, data in the format of numbers and / or text, etc. or agronomical input data suitable for re-training a pre-trained transformer-based model or using a trained transformer-based model for agronomical resources. Thus, the agronomical environment should be construed broadly as a technical environment producing an agronomical product, a physical product and / or a service associated with a product; a product may even be a data product, an answer to a query or an advice and while producing said product generating, by said agronomical environment, multi-parameter agronomical related data flows and / or agronomical data.

[0074] FIG. 1 illustrates an agronomical environment comprising one or more agronomical resources. In an example the agronomical resources may be allocated to a specific user and thus may be user specific.

[0075] An agronomical environment may comprise equipment 1002 and sensors 1003 generating one or more sensor related data flows. The agronomical environment may produce one or more agronomical products, wherein properties of said one or more agronomical products may be measured, extracted or calculated generating one or more agronomical product related data flows. Agronomical data 1005 may comprise data obtained from each of said one or more data flows. Agronomical data 1005 may also comprise description data. Description data may comprise description data of at least one tool and / or of at least one repository.

[0076] Equipment 1002 may be any equipment of an agronomical environment such as actors, pumps, nozzles, valves, reaction tanks, separation chambers and / or alike.

[0077] Sensors 1003 may be any kind of sensors of an agronomical environment such as temperature sensor, weather sensor, flow sensor, pressure sensor and / or alike.

[0078] One or more agronomical products produced by the agronomical environments may be any type of products as described above. Properties of the agronomical products and / or the agronomical resource may be measured by, for example, balance.

[0079] Agronomical data may also be instructions on how to react to certain defects. In an example, how to improve the performance of a field. Agronomical data 1005 may be stored in a database or repository, e.g., as historic data.

[0080] Agronomical data 1005 may be provided to the pre-processing engine that pre-process the data and provides agronomical input data to an analytics engine 1004 that may analyze the agronomical input data and generate machine readable instructions 1007 for control and / or monitoring engine 1006. Generation of the machine- readable instructions 1007 may be automatic (i.e., without involving an operator or user) by the analytics engine 1004 based on the analysis of the input agronomical data. For example, the analytics engine 1004 may continuously receive agronomical input data and may analyze said data in a continuous mode.

[0081] When a solution and / or answer to an inquiry occurs, the analytics engine may generate machine readable instructions 1007 for the control and / or monitoring engine to provide an answer to the inquiry. The analytics engine may identify solutions for improving agronomical production efficiency by analyzing agronomical input data, e.g. a question of the user, on the background of the agronomical task or production processes and may send machine readable instructions 1007 to the control and / or monitoring engine for improving production, the agronomical task and / or the situation of the agronomical resource.

[0082] The control and / or monitoring engine 1006 may display a push notification to a user to review the instructions 1007 and based on the review may generate reviewed machine-readable instructions 1008. The control system, based on the machine-readable instructions 1008, may change the operating parameters of one or more pieces of equipment 1002. The machine-readable instruction 1007 and / or reviewed machine-readable instruction 1008 may also comprise information about an agronomical source and an answer to an agronomical question provided by the user. Reviewing said machine readable instructions may allow for safe integration of the trained transformer based model in the agronomical environment.

[0083] Alternatively, or in addition to the automatic generation of machine-readable instructions 1007 by the analytics engine, a user and / or an assistance device may prompt the analytics engine 1004 to provide machine-readable instruction 1007 based on a prompt as described, for example, in the context of FIG. 7. The prompt provided by the user may comprise an agronomical question. An assistance device may receive query data from a user and forward the query data to the trained transformer base model.

[0084] Machine readable instructions 1007 may be used by a user for controlling and / or monitoring of one or more agronomical tasks of the agronomical environment. Machine readable instructions 1007 may comprise operating instructions for the agronomical resource such as machine-readable instructions for controlling equipment 1002 and / or machine-readable instructions for monitoring equipment 1002 and / or sensors 1003. Machine readable instructions 1007 may comprise operating instructions for a user for controlling and / or monitoring the agronomical environment, wherein the operator may be a human based operator, computer-based operating system or a hybrid system comprising a human operator and a computer-based assistance system. The machine-readable instructions 1007, 1008 may comprise information about an agronomical resource. They may comprise instructions of how to improve the performance of a field. The machine-readable instructions may be provided in natural language form. They may be convertible into speech or audio output and / or data associated with natural language.

[0085] A user may review the machine-readable instructions 1007 and, based on the review, may generate machine readable instructions 1008 for controlling equipment 1002 and / or sensors 1003. A user may prompt the analytics engine 1004 to provide operating instructions (machine-readable instructions 1007) based on a prompt, a query, a context and / or alike as illustrated in FIG. 7. The query may comprise agronomical querying data associated with an agronomical query from a user.

[0086] A user may be a human operator reviewing the machine-readable instructions 1007. A user may be a human operator having a computer-based assisted system for reviewing the machine-readable instructions 1007. A user reviewing the machine-readable instructions 1007 may be a computer-based system based on a computer program comprising a set of instructions for reviewing machine readable instructions 1007.

[0087] The control system of the agronomical environment may comprise one or more computing units that may be able to manipulate one or more process parameters related to the agronomical task by controlling one or more of the actuators or switches and / or end effector units, for example via manipulating one or more of the equipment operating conditions. The controlling is typically done in response to the one or more signals retrieved from the equipment.

[0088] The control and monitoring engine may comprise one or more computer processors for revising machine-readable instructions 1007 and generating reviewed machine-readable instructions 1008 for the control system of the agronomical production. The control system of the agronomical production, based on the machine-readable instructions 1008, may adjust the equipment operating conditions such that the adjusted process parameters and / or equipment operating conditions result in a controlled product (such as an agronomical product) that has one or more required or pre-determined properties or performance parameters. The agronomical product may also comprise instructions on how to treat an agronomical resource in order to receive a desired result. An agronomical production can thus be controlled on-the-fly whilst ensuring that the equipment operating conditions are adapted to undesired variations in the process parameters of the agronomical tasks.

[0089] It may be understood that controlling and monitoring of an agronomical environment in general relates to controlling equipment, an agronomical resource and / or to treat an agronomical resource for producing an agronomical product by sending machine readable instructions to the agronomical environment and / or to the user. An agronomical product should be construed broadly as described above. As a further example, an agronomical product may even be a data product, and / or a data service product provided by the trained transformer-based model trained on agronomical training data based on the agronomical resource.

[0090] In an example the agronomical training data and / or the agronomical data comprise description data, e.g. an API description, of at least one tool and / or of at least one repository. The at least one tool and / or on the at least one repository is adapted to provide at least one functionality and / or at least one information unit selected from the group of functionalities and / or information units, including structured information, unstructured information, public information, private information, a geospatial information, map data, sensor data, domain specific information, agronomic knowledge, regulation information, mathematical functionality, yield price retrieving functionality, encyclopedic functionality (Wikipedia, SPARQL), and shipping allocation functionality.

[0091] The agronomical training data may also comprise interface description of available resources and / or tools and in this way, training may comprise teaching the pre-trained transformer-based model with specific knowledge about available information.

[0092] The transformer-based model (first ML model) operated by the Al engine as described in the context of FIGs. 2a, 2b, 9-12 (the first ML model) may be integrated via an interface, in particular a computer interface (e.g., an API, a GUI, a web application) with another computer program, such as a second ML model, wherein the second ML may comprise a different algorithm (e.g., a classical ML not based on a transformer architecture, a variation of the transformer based architecture of the first model or alike) and / or the second ML may comprise the same architecture as the first model but the second model may be trained on a different dataset.

[0093] The second ML model may be used for pre-processing of raw agronomical data to generate agronomical training data. Pre-processing of the raw data may involve removing noise, filtering data, labeling data, sorting data, converting raw data into a different format, converting operator data into a different format better suitable for the first ML model (transformer based model), and or alike. Said pre-processing of the raw data may also be done by a computer program based on a set of computers implemented instructions (e.g., a set of programming instructions, filters, labels, mathematical steps and / or alike not comprising a ML model). Pre-processing of raw agronomical data may improve the quality of training or input data based on the raw data and may reduce the computing power required for training / using the first ML model (transformer-based model).

[0094] FIG. 2A illustrates an operating system of the agronomical environment of FIG. 1. The operating system comprises training and / or re-training a pre-trained transformer-based model and using a trained transformerbased model for controlling and / or monitoring an agronomical resource such as a field. Training a pre-trained transformer-based model is further described in the context of FIG: 2B. Using the purpose trained transformerbased model is further described in the context of FIG. 2C.

[0095] FIG. 2B illustrates training a pre-trained transformer-based model on agronomical data generated by the agronomical environment of FIG. 1 .

[0096] Training, fine-tunning and / or re-training of a pre-trained transformer-based model, such as described in the context of FIGs. 9-12, comprises accessing a pretrained transformer-based model via an interface, in particular a computer interface (GUI, API, web-based interface) and receiving, via the interface, historic agronomic data 1010 from pre-processing engine 1009. A user and / or assisting device re-training the pre-trained transformer-based model may prompt the pre-trained transformer-based model to access the historic agronomic data via the computer interface, or alternatively, the user may upload, via the interface, in particular the computer interface, the historic agronomic data 1010 from a database to the Al engine that comprises at least one processor used for operating the pre-trained transformer-based model.

[0097] The user re-training the pre-trained model may be a human user and / or an assistance device, an automated operating system comprising a computer processor, or a hybrid operating system comprising a human user and a computer assisted operating system prompting the pre-trained model to train or re-train on agronomical data.

[0098] The historic agronomic data 1010 may be based on agronomic data 1005. The historic agronomic data 1010 may comprise sensor data, description data of the at least one tool and / or of the at least one repository and / or documents comprising agronomical knowledge.

[0099] Agronomical data 1005 and historic agronomical data 1010 may be stored in a database or repository. Agronomical data 1005 may comprise a plurality of data of the agronomical resource as described, for example, in the context of FIG. 4.

[0100] During the training, the historic agronomical data 1010 may be embedded via an embedding layer as described within the context of FIG. 9. Embedding the agronomical data may result in embedded agronomical data and / or description data of the at least one tool and / or of the at least one repository offered in an agronomical environment.

[0101] The above examples of agronomical data are mere examples. Agronomical data should be construed broadly and should be understood as any kind of data associated with producing an agronomical product using the agronomical resource. Any type or format of data associated with production and / or with the agronomical task may be pre-processed by the pre-processing engine into data suitable for training or using the transformer based model for production.

[0102] At the end of a training cycle, the analytics engine 1004 may output and / or release a trained transformer-based model suitable for use in an agronomical environment. The released trained model may be stored in a database for, e.g., version control of released models. The released trained model may be a computer program product. An access to the released trained model may be provided to a user and / or to an assistant device as a data service for assisting production.

[0103] Training agronomical data may be agronomical data in one or more languages. Training the pre-trained transformer-based models on training agronomical data in one or more languages allows for enlarging agronomical training dataset, and additionally, allows for providing operating instructions for an agronomical task and / or for production in one or more languages.

[0104] Providing instructions in one or more languages may improve user interaction with the trained transformer-based model.

[0105] FIG. 2C illustrates an embodiment of using a trained transformer-based model for controlling the agronomical environment shown in FIG. 1.

[0106] The agronomical data 1005 may be provided to the pre-processing engine 1004 for pre-processing. Preprocessing the agronomical data 1005 may comprise steps as, for example, described in the context of FIG. 3. Pre-processing may also involve other steps such as labeling data, removing noise, structuring unstructured data, converting data into different formats, and / or alike required to provide training / input data based on agronomical data suitable for training or using a transformer-based model for production.

[0107] The pre-processed data from the pre-processing engine 1004 may be used as agronomical input data for the trained transformer-based model for an agronomical task. The agronomical task may comprise identifying properties of an agronomical resource and / or identifying appropriate measures for the agronomical resource in order to change the properties. In an example the property may be a performance of a field.

[0108] The trained transformer-based model may receive the agronomical input data and may, for example, predict solution and / or answer to an inquiry for the agronomical input data 1011 , e.g. for agronomical querying data associated with an agronomical query of a user. Based on analyzing the agronomical input data 1011 , the trained transformer-based model may, for example, predict how to resolve errors and / or how to reach a desired result in an agronomical task, how to improve an agronomical property of a resource, how to resolve user queries associated with controlling and / or monitoring of the agronomical resource and alike. Based on the perdition, the analytics engine 1004 may generate machine readable instructions 1007 for control and / or monitoring engine 1006. The instructions may include links to resources and / or tools which may support solving the agronomical task.

[0109] The analytics engine 1004 operating the pre-trained transformer based model may have at least one interface, in particular a computer interface, e.g., graphical user interface, GUI, a web based computer interface, and / or application programming interface, API, for operating the transformer based model. The operation of the transformer-based model may comprise uploading data, providing prompts such as prompts comprising instructions for accessing training data based on the agronomical resource and / or agronomical input data, providing prompts for re-training, providing prompts for generating machine readable instructions 1007, means for reviewing machine readable instructions 1007, means for receiving notifications when new machine readable instructions are generated, and alike. The model and / or a copy of the model, either untrained, pre-trained and / or trained, may be stored in a database, on a server or in a cloud. The model may be operated by the Al engine via the at least one computer interface, e.g., graphical user interface, GUI, a web based computer interface, and / or application programming interface, API . In an example the model is run in a private cloud and fine-tuned with additional knowledge, e.g. special agronomical knowledge.

[0110] A user of the trained transformer-based model may provide an input via a computer interface, e.g., a graphical user interface, application programming interface, a web-based interface, to the trained transformer-based model such as a text and / or audio query as, for example, described in the context of FIG. 5. The user of the trained transformer-based model may additionally provide extracted data points, extracted, for example, from the data provided by the pre-processing engine. The extracted data points may, for example, relate to a solution and / or answer to an inquiry for the agronomical input data, e.g. agronomical querying data, and are associated with data points, typical or optimal parameters of an agronomical task of the agronomical environment. The extracted data points together with the user’s query may form a part of the input data to the trained transformer-based model. The trained transformer-based model may process the input data 1011 and provide a solution to the user query, for example, in the form of machine-readable instructions 1007.

[0111] The control and / or monitoring engine may review the machine-readable instructions 1007.

[0112] After the review, the control and / or monitoring engine 1006 may generate machine readable instructions 1008 that may be identical to machine readable instructions 1007, may be in part based on machine readable instructions 1007 or may be different from machine readable instructions 1007. A user of the control and / or monitoring engine

[0113] 1006 may use machine readable instructions 1007 merely for monitoring the agronomical task and / or the agronomical resource or may forward machine readable instructions 1007 as machine readable instructions 1008 for controlling the agronomical task and / or the agronomical resource. The user may be a human user, an assisting device and / or an operating system comprising a processor.

[0114] In case when machine readable instructions 1008 generated by the control and / or monitoring engine are different from machine-readable instructions 1007 generated by the trained transformer based model, the control and / or monitoring engine may trigger re-training of the trained transformer-based model. The re-training may be triggered automatically based on the feedback from the control and / or monitoring engine, the feedback may indicate that the machine-readable instructions 1007 and 1008 deviate. The feedback may comprise a feedback score indicating the degree of deviation.

[0115] Re-training may be triggered by a user. Alternatively, or additionally, re- training may be scheduled or continuous.

[0116] The trained transformer-based model may be trained on agronomical training data in one or more languages. The trained transformer-based model may provide operating instructions for an agronomical task in one or more languages. Providing operating instructions in one or more languages may be advantageous, for example, in that training data may be more available in one language than in another language. The trained transformer-based model trained on agronomical training data in one or more languages may be set to provide operating instructions

[0117] 1007 in a language in which the most training data is available. Additionally, or alternatively, the trained transformer-based model may be set to provide operating instructions 1007 for controlling and / or monitoring an agronomical resource 1001 in a language of choice of a user making the operation of the model more user friendly. Additionally, or alternatively, the trained transformer-based model may be requested to provide operating instructions for controlling and / or monitoring an agronomical resource 1001 in more than one language for cross checking operating instructions and selecting the most suitable instructions for improved production. In yet another example the operating instructions may be provided as audio or speech signal in natural language. FIG. 3 illustrates a pre-processing engine 1009 of the operating system as shown in FIG. 2.

[0118] Raw agronomical data 1005 and / or description data as described in the context of FIG: 4 may be provided to preprocessing engine 1009.

[0119] The pre-processing engine 1009 may pre-process raw agronomical data to provide agronomical training data and / or agronomical input data to the transformer-based model. Pre-processing steps may comprise selecting required parameters, merging / aggregating, calculating agronomical training data, e.g., calculating derived parameters, remove outliers and alike. Pre-processing may comprise filtering the data, removing noise, labeling of data, sorting data, converting data from formats that are not suitable for training / using the model into formats that are suitable for training / using the model and / or alike. The output data of the pre-processing engine 1009 may be stored in a database and used as agronomical training data for re-training a pre-trained a transformer-based model as described in the context of FIGs. 2a, 2c, 6 and 9 to 12. The output data of the pre-processing engine may also be used as input data for the trained transformer-based model as described in the context of FIGs. 1 , 2a, 2c and 7.

[0120] FIG. 4 illustrates agronomical data structure from agronomical resource(s) as shown in FIG. 1 .

[0121] Agronomical data may be received from the agronomical environment via a computer interface (e.g., a graphical user interface, an application programming interface, a web-based interface).

[0122] The agronomical data may comprise different categories of agronomical data, e.g., sensor data, operating data, agronomical metadata, description data.

[0123] Sensor data may relate to measured quantities available in an agronomical environment by means of installed sensors, e.g. temperature sensors, pressure sensors, flow rate sensors. LIDAR sensors, cameras, etc.

[0124] Description data may relate to description data of an API of the at least one tool and / or of the at least one repository. The description data may refer to locations of available data usable in combination with the trained transformer-based model. By retraining the trained transformer-based model may be fine-tuned in such a way that the trained transformer-based model learns about additional data sources usable for responding to queries and / or to find solutions. The API data may also comprise an address and / or endpoint, to indicate how the corresponding additional tool and / or repository may be reached.

[0125] Operating data may relate to raw data, e.g. basic, non-processed analytical and / or sensor data, or processed or derived parameters, directly or indirectly derived from raw data. Agronomical metadata may comprise scientific paper and describe agronomical strategies, e.g. how to treat an underperforming field in order to increase the performance.

[0126] The agronomic data 1005 may comprise text and / or numbers as structured data and / or unstructured data. Agronomical data and / or agronomical metadata may particularly be available in unstructured format. The unstructured data may comprise scans, documents, datasheets comprising images, QR codes, and alike, may be pre-processed by the pre-processing engine and converted into text and / or numbers suitable for training of a pretrained transformer-based model as described in the context of FIGS. 9 to 12.

[0127] Each category of data may be provided in a structured or unstructured format and / or as private or public data.

[0128] FIG. 5 illustrates data structure from a user.

[0129] A user, e.g., an operator of an agronomical environment and / or a farmer having at least one field, may input a query to the analytics engine as, for example, text or audio query regarding monitoring and control of distributed production or of a distributed agronomical task. The query may be, for example, a request for providing operating instructions 1007 for improving the performance of a field and / or a query to show the 5 most performant fields. For example, the user may request the trained transformer-based model to provide operating instructions for resolving an agronomical task, e.g. improve performance of a field, resolving an error on an agronomical resource. The analytics engine may predict a solution to the query based on the input from the user and the agronomical input data. The solution may comprise machine readable instructions 1007, e.g. operating instructions 1007, for example, for increasing dosage for equipment 1002, e.g. as an application map for an application device.

[0130] FIG. 6 illustrates further details of re-training a pre-trained transformer-based model for an agronomical environment in addition to embodiments in FIGs. 2a and 2b.

[0131] A pre-trained transformer-based model pre-trained on text and / or numbers (such as illustrated in FIGs. 9 to 12) may be trained with agronomical historic data (agronomical training data). In particular, the transformer-based model may be trained and / or parametrized as described within the context of FIG. 2b and FIG. 1 1 , training and / or deploying the transformer encoder, the transformer decoder and / or the transformer encoder-decoder. Training steps as described in context of FIGs. 9-12 may be followed for re-training the model for an agronomical task using agronomical data instead of generic text / numbers as illustrated in FIGS 9 to 12. In an example the training data comprise description data, e.g. API description, of the at least one tool and / or of the at least one repository. A trainer of a pre-trained transformer-based model may create a prompt to trigger re-training of the pre-trained transformer-based model and indicate or upload agronomical training data based on agronomical data 1005 for training, e.g., via a computer interface, e.g., web based interface, API, GUI. At the end of the training cycle, the trainer of the model may test the trained model, and based on the testing, the trainer may run additional training cycles or release the trained model for use in an agronomical environment as described in the context of FIGs. 2a and 2c.

[0132] The released trained transformer-based model may be stored in a database or in a cloud. The released model may be operated by a computing unit / node comprising a computer processor such as the analytics engine. A copy of the trained transformer-based model may be provided to a user for use and / or further training. Alternatively, only an access to operate the trained transformer-based model may be provided to a user.

[0133] Instead of the pre-trained transformer-based model as described in the context of FIGs. 9 to 12, another foundation model may be used by the analytics engine 1004, such as, for example, ChatGPT (GPT-2, GPT-3, GPT-3-5-turbo, GPT-4 or higher / similar), Davinci, BERT (Bidirectional Encoder Representations from Transformers), DistilBERT, Transformer-XL, XLNet (extreme Language understanding Network), T5 (Text-to-Text Transfer Transformer), RoBERTa (Robustly Optimized BERT approach), ELECTRA (Efficiently Learning an Encoder that Classifies Token Replacements Accurately), Reformer, Longformer, DeBERTa (Decoding-enhanced BERT with disentangled attention) and alike, or any other large language model pre-trained on big data such as generic text, image, video.

[0134] Depending on availability of agronomical training data, availability of computing resources and required precision in analyzing data, a user may choose a foundation model with smaller or larger number of parameters. The transformer-based model shown in FIGs. 9 to 12 may be pre-trained to release a pre-trained model with the required number of parameters to suit the technical purpose of the user. The pre-trained model pre-trained in the context of FIGs. 2b, 9 to 12 may be further refined to release a model with even less parameters for improved computing speed and reduced computing resources.

[0135] FIG. 7 illustrates contextualization of prompts.

[0136] A user, e.g. a user of the agronomical environment, may prompt, via a computer interface, e.g., a graphical user interface, an application programming interface, a web-based interface, a trained transformer-based model as described in the context of FIGS. 2a, 2b and 6 to resolve a user query related to how to control / monitor an agronomical resource. The user may further provide a context of the query, for example, sensor data, operating data, agronomical metadata, description data. The user may select the context via the computer interface or enter the context via text and / or audio channel in one or more languages. The user may provide a second context. The second context may, for example, comprise one or more key words (e.g., answer to a query, error associated with a piece of equipment X, error in software operating a piece of equipment X, answer to a query including content of at least one tool and / or of at least one repository referenced by description data), one or more datapoints from agronomical data, e.g., one or more data points relating to a query, one or more standard operating data and alike.

[0137] Based on one or more contexts, and optionally, agronomical input data, the user may prompt the trained transformer based model to predict a solution to the user query. In particular a query with regard to the resource. The predicted solution may comprise machine readable instructions 1007 sent to control and / or monitoring engine

[0138] 1006 for operating the agronomical environment 1001 .

[0139] A user may provide one or more contexts in one or more languages and prompt the trained transformer-based model to provide operating instructions for an agronomical task in one or more languages.

[0140] FIG. 8 illustrates trigger-based re-training of a trained transformer-based model when the model is in use for an agronomical task.

[0141] Once trained as described in the context of FIG. 2b, 6 and FIG. 1 1 , the trained transformer-based model may be further re-trained or fine-tuned based on a feedback related to the operating instructions provided by the trained transformer-based model. For example, a user and / or an operator may monitor machine readable instructions

[0142] 1007 provided by the trained transformer-based model to the control and / or monitoring engine 1006. The user may further assign a feedback score and a threshold related to the operating instructions provided by the trained transformer-based model. In the event of the score falling below a threshold value, re-training may be triggered automatically by a computer processor (e.g., the control / monitoring engine or the analytics engine), or the training may be triggered by a human operator. For example, the score may be below the threshold value when the machine-readable instructions 1007 are not acceptable by the operator for controlling the agronomical equipment 1002 (e.g., in case the machine-readable instructions 1007 are contradictory to optimal operating conditions of said equipment 1002).

[0143] The re-training cycle may comprise steps as illustrated in the context of FIG. 2b, 6 and FIG. 9-11 , wherein the training data are agronomical data. After a re-training cycle, the user or operator may provide a new feedback score for new machine-readable instructions 1007. The training may continue until the feedback score reaches the threshold or a maximum number of training cycles. At the end of re-training, the model may be stored and / or released for use in an agronomical task. Re-training may be carried out on the background of ongoing production without the need to stop / interrupt the agronomical task.

[0144] FIG. 9 illustrates an embodiment of training an embedding layer. The embedding layer may be obtained by training for example a continuous bag of words model (CBOW) or a skip-gram model. The embedding layer may be suitable for generating embedded input data based on input data. Generating embedded input data may refer to embedding input data. Embedding input data may result in a representation associated with the input data. Thus, the embedded input 114 may be the representation associated with the input data. The input data may comprise a one or more elements. The one or more elements may be represented by the input vector 106. In particular, the embedded input 1 14 and / or the input vector 106 may be machine- readable and / or processable by a processor. For this purpose, the embedded input 114 and / or the input vector 106 may be a tensor, in particular a first-rank tensor. Specifically, the input vector 106 may be a one hot vector or a summation of a plurality of one hot vectors. A one hot vector may be a vector with one entry unequal to zero. Examples for one hot vectors may be 108, 1 10 and 112. The entries unequal to zero in the one hot vector and / or in the input vector 106 may indicate the element. For example, a look up table may define the relation between the position of the entries unequal to zero and the element indicated by the one hot vector.

[0145] The look up table may specify a plurality of different elements. The number of different elements may be equal to the number of entries in the one hot vector. The number of different elements may be referred to as vocabulary size. In an example, the elements may be represented by tokens and a sequence of elements may refer to at least a part of a sentence.

[0146] The at least a part of the sentence may be represented by a plurality of tokens. A token may represent at least a part of the element and / or word. For example, where one element would be associated with only one word, words such as “embeddings”, “embedding” or “embed” would constitute different elements. A first token may represent the stem “embed” and the endings, typically appearing in a plurality of word, may be represented by a second token, a third token and a fourth token. The second token, the third token and the fourth token may be used for representing other words such as “look”, “looking” or the like, preferably together with a fifth token representing the stem “look”. Ultimately, this tokenization of elements associated with a plurality of stems and a plurality of endings results in less tokens to be used for representing a plurality of elements and thus, uses less computational resources.

[0147] A look up table specifying a subset of the vocabulary size eg of the English language may comprise 10,000 words or more. The embedded input 114 may be a lower-dimensional representation than the input vector 106. For example, typical embedded inputs 114 may comprise some hundreds of different entries. Followingly, the embedded inputs 114 constitute a densified representation of one or more elements using less computational resources. More than that, the embedded input 114 may represent a relation between two or more elements. For example, the words “Italy” and “Germany” may be similar or may be more closely related since they both define European countries, whereas the word “embodiment” may be very different from the two respective words. The smaller the dot product between two embedded inputs 114 may be the more similar the two elements associated with the embedded inputs 1 14 may be. Hence, the embedded inputs 114 may represent one or more elements accurately and lead to accurate results based on processing the embedded inputs 1 14.

[0148] For transforming the input vector 106 into the embedded input 114, the embedding layer may comprise a number of neurons equal to the number of entries in the embedded input 1 14.

[0149] Based on the embedded inputs 114, the output layer may generate the output vector 116. The output vector may be a vector and / or may indicate one or more elements. The output vector 1 16 may indicate one or more elements different from the input vector 106 and / or the one hot vectors associated with the input vector 106. For this purpose, the output layer may comprise a number of neurons equal to the number of entries of the input vector 106 and / or the output vector 1 16. The output layer may apply a softmax function to the embedded inputs 1 14. By doing so, the output vector may comprise the probabilities associated with the elements associated with the entries of the output vector 1 16 unequal to zero. Hence, from the output vector 1 16 one or more elements may be obtained with a corresponding probability. Where the input vector 106 may specify one or more sequence(s) of elements, the output vector 1 16 may specify one or more elements corresponding to the sequence(s) of elements specified by the input vector 106. In the example of FIG. 9, the element associated with vector 118 may correspond to the input vector with a probability of 71 %. Additional or alternative elements may correspond to the input vector as indicated by the output vector with lower probability. By defining a threshold to which the probability may be compared, the selection of the corresponding elements may be tailored to the needs of the user. The elements generated by the model comprising the embedding layer 102 and the output layer 104 may refer to the most probable elements indicated by the output vector 1 16. Hence, the model depicted in FIG. 9 may generate the element associated with the vector 118 with a confidence score of 71 %.

[0150] The model of FIG. 9 may be continuous bag of words (CBOW) model. The CBOW model may be trained based on a training data set comprising a plurality of input vectors and corresponding output vectors. As the training data set may not be labeled, the training of the CBOW model may be referred to as self-supervised. Before training of the CBOW model, the CBOW model may be initialized with random values assigned to the weights of the neurons.

[0151] During the training of the CBOW model, the input vectors may be passed through the initialized embedding layer and the output layer and a loss may be determined by comparing the output vector obtained by passing the input vector 106 through the model to the output vector corresponding to the input vector 106 as specified by the training data set. Based on the determined loss, backpropagation may be applied to determine the gradients associated with the neurons of the embedding layer 102 and the output layer 104 to lower the loss.

[0152] According to the determined gradients, the weights of the neurons may be updated by using a gradient descent algorithm. If a predetermined loss may be achieved by the CBOW model, the training may be terminated and a trained CBOW model may be obtained. From the trained CBOW model, the embedding layer 102 may be suitable for embedding input data comprising one or more elements. This embedding layer 102 may be used in other machine-learning architectures requiring an embedding layer 102 such as a transformer encoder, transformer decoder or transformer encoder decoder architecture as described within the context of FIG. 10A, FIG. 10B and FIG. 10C. For training these architectures, a trained embedding layer 102 may be required. Hence, a model such as a CBOW model may be trained prior to training the transformer encoder, transformer decoder or transformer encoder decoder architecture.

[0153] FIG. 10A illustrates an embodiment of a transformer encoder architecture. The transformer encoder comprises an encoder input 278, one or more encoder blocks 274, 214 and an encoder output. The transformer encoder architecture may be derived from the transformer encoder-decoder architecture as known in the art and shown in FIG. 10C. In particular, the transformer encoder may be referred to as X-former. The transformer encoder architecture may correspond to the encoder architecture associated with the transformer encoder-decoder architecture with an additional encoder output instead of connecting the encoder block directly to the decoder of the transformer encoder-decoder architecture. A plurality of transformer encoder architectures are available in the art such as the bi-directional encoder representations from transformers (BERT).

[0154] The input data may be received at the encoder input 278. The encoder input 278 may apply an input embedding 202. Applying the input embedding 202 may refer to passing the input data through an embedding layer eg as described within the context of FIG. 9. Further, the encoder input 278 may apply positional encoding 204. Applying positional encoding 204 may refer to adding a positional factor to the embedded input obtained via input embedding. Preferably, the input data may specify a sequence of elements. The positional factor pposmay be indicative of the position of the elements within the sequence. For example, the positional factor pposmay be obtained based on the following equation: where pos may refer to the position of the element within the sequence, I may refer to the dimension associated with the input embedding and d may refer to the dimension of the model, e.g. transformer decoder, transformer encoder or transformer encoder-decoder. This may be referred to as absolute positional embeddings.

[0155] Alternatively, the positional encoding may be based on rotary positional embeddings (RoPE). Positional encoding is beneficial since it enables the processing of sequential data without requiring further dimensions indicating the position of each element. Followingly, the positional encoding 204 reduces the computational resources needed for embedding the input data. By passing the input data through the encoder input, the input data may be transformed into a second-rank tensor representing the sequence of elements. This second-rank tensor may be referred to as embedded input data.

[0156] The embedded input data may be processed by the encoder block. The embedded input data may be provided to the layer normalization 208 by a residual connection. Multi-head self attention 206 may be applied to the embedded input data. Multi-head self attention 206 may comprise the two components multi-head and selfattention. Self-attention may be understood as being a filter applied to the embedded input data. By applying the filter to the embedded input data, the elements associated with the embedded input data contributing to the to be generated output data may be identified for generating the output data. Hence, the filter may represent the degree of contributing to the to be generated output data by the elements associated with the embedded input data. Applying the filter may be referred to as weighting the elements associated with the embedded input data. This is advantageous specifically regarding long sequences of elements. The filter may be learned and improved during the training by learning to identify the contribution of elements associated with the embedded input data. For example, in the partial sentence “I went to the bakery to buy a” the last word may be generated by the data-driven model such as the transformer encoder. The self attention may focus the transformer encoder to attend to the word “bakery” and “buy” mostly to generate the word “bread”. Self attention may refer to attention generated based on the input data. Hence, the filter may be determined based on the input data, preferably the embedded input data. The embedded input data may serve as query Q, key K and value V with respect to the self attention operation. The self attention may refer to attention based on the received input data.

[0157] Hence, the filter may be calculated based on the following formula by inserting the respective tensors based on the embedded input data: corresponds to the dimension of the key. For improving the efficiency of the transformer encoder further, the multiple heads are used to apply the filter resulting in the multi-head self attention 206. Multi-head self attention 206 may comprise applying the filter to two or more parts of the embedded input data. Hence, the tensor may be split into two or more parts and the filter may be applied to the two or more parts separately by two or more heads according to the following equation:head i =Attention QWiQ, KWiK, VWiV) with parameter matrices j .Qe^dxdQ, WiKE ^dxd^ jyT£jjdxdy where I may refer to the number of heads, may refer to the dimensions of the value, key and query.

[0158] The result of the two or more head may be concatenated according to the following equation: MultiHead(Q, K, V) = Concat (head. 1, . . . , headh( W'’ where yye^hdv*d and h may refer to the number of heads.

[0159] The embedded input data may be transformed via the multi-head self attention 206 into a context tensor. The context tensor may represent the sequence of elements and the relation between two or more elements of the input data. The context tensor may be a second rank tensor and / or may comprise one or more first rank tensor(s). After the multi-head self attention 206 layer normalization 208 may be applied based on the context tensor and / or the embedded input data from the residual connection. Applying layer normalization 208 may refer to normalizing the context tensor. Normalizing the context tensor may lower the values of the entries of the context tensor. This reduces the computational cost associated with processing the context tensor. Layer normalization 208 may be followed by passing the context tensor to a feed forward layer 210 again followed by layer normalization 212 based on the residual connection to the context tensor and / or the output of the feed forward layer 210. The feed forward layer 210 may be a feed-forward neural network. The feed-forward neural network may comprise of a plurality of fully connected neurons. Passing the context tensor through the feed-forward neural network may result in transforming the context tensor linearly. Additionally or alternatively, the neural network may comprise one or more activation functions such as a rectified linear unit (ReLU). Hence, the neural network may be configured for performing one or more non-linear operations to the context tensor and / or transforming the context tensor non-l inearly . After the context tensor has been transformed and / or normalized by the feed forward layer 210 and the layer normalization 212, the context tensor may be provided to one or more further encoder blocks 214. Having passed the context tensor through the feed forward layer 210 may adapt the context tensor for the processing by a further attention layer of the one or more further encoder blocks 214 for applying a self attention filter, preferably multi-head self attention 206. The context vector after being transformed by the layer normalization 212 and the feed forward layer 210 may be referred to as hidden state. The encoder output 276 comprises of a linear layer 216 and a softmax layer 218. The linear layer 216 may transform the context vector into a logits vector. The linear layer may be fully-connected. The logits vector obtained by passing the context tensor through the linear layer 216 may be passed through the softmax layer 218. Passing the logits vector through the softmax layer 218 may refer to applying the softmax function to the logits vector. Applying the softmax function to the logits vector may result in a probability distribution of one or more elements corresponding to the sequence of elements in the input data. From the probability distribution based on predefined selection criteria, one or more elements may be chosen. The one or more chosen elements may be referred to as the one or more elements generated by the transformer encoder. The one or more generated elements may be provided to the encoder input for generating further one or more elements corresponding to the sequence of the input data and the one or more elements generated by the transformer encoder as described within the context of FIG. 11 .

[0160] FIG. 10B illustrates an embodiment of a transformer decoder architecture.

[0161] The transformer decoder comprises a decoder input 284, one or more decoder blocks 280, 232 and a decoder output 292. The transformer decoder architecture may be derived from the transformer encoder-decoder architecture as known in the art and shown in FIG. 10C. The transformer decoder may be referred to as X-former. The transformer decoder architecture may correspond to the decoder architecture associated with the transformer encoder-decoder architecture independent of receiving one or more hidden states from the encoder of the transformer encoder-decoder. A plurality of transformer decoder architectures are available in the art such as the generalized pretrained transformers (GPT).

[0162] The decoder input 284 may apply input embedding 220 and positional encoding 222 analogous to analogous to the input embedding 202 and the positional encoding 204 as described within the context of FIG. 10A.

[0163] The decoder block 280 may comprise the layer normalizations 226, the masked multi-head self attention 224, the feed forward layers 228 and / or the layer normalization 230. The embedded input data resulting from passing the input data through the decoder input 284 may be provided to the layer normalization 226 via a residual connection. Further, masked multi-head self attention 224 may be applied to the embedded input data. Masked multi-head self attention 224 corresponds to the multi-head self attention 206 as described within the context of FIG. 10A with additionally masking a part of the embedded input data associated with elements later in the sequence than the element to be generated. Additionally or alternatively, the part of the input data associated with elements later in the sequence than the element to be generated may not be received and / or transformed into the embedded input data. Thus, the transformer decoder may be suitable for generating a subsequent element to a sequence, whereas the transformer encoder may be suitable for generating a missing element in within one sequence and / or between two or more sequences. Therefore, the transformer encoder may be configured for classification tasks. The transformer decoder may be configured for text generation.

[0164] Similar to the transformer encoder as described within the context of FIG. 10A, a context tensor may be generated by applying the masked multi-head self attention 224 and the layer normalization 226. The context tensor may be provided to the layer normalization 230 via a residual connection. Further, the feed forward layer 228 and the layer normalization 230 may be analogous to the feed forward layer 210 and the layer normalization 212 as described within the context of FIG. 10A. The context tensor may be provided to one or more further decoder blocks 232.

[0165] The decoder output 292 may comprise of a linear layer 234 and a softmax layer 236. The linear layer 234 and the softmax layer 236 may be analogous to the linear layer 216 and the softmax layer 218 as described within the context of FIG. 10A.

[0166] FIG. 10C illustrates an embodiment of a transformer encoder-decoder architecture. The transformer encoderdecoder may comprise the encoder input 288, the one or more encoder blocks 286, 264, the decoder input 294, the decoder block 290 and the decoder output 292.

[0167] The encoder input 288 may correspond to the encoder input 278 of FIG. 10A. The one or more encoder block 286, 264 may correspond to the one or more encoder blocks 274, 214 of FIG.

[0168] 10A. The decoder input 294 may correspond to the decoder input 284 of FIG. 10B.

[0169] The decoder block 290 may comprise a masked multi-head self attention 270, a layer normalization 272, a feed forward layer 238 and a layer normalization 240 analogous to the masked multi-head self attention 224, the layer normalization 226, the feed forward layer 228 and the layer normalization 230 as described within the context of FIG. 2B. The decoder block 290 may further comprise a multi-head self attention 250 and a layer normalization 248.

[0170] Analogous to the description of FIG. 10B, the context tensor may be obtained from the masked multi-head self attention 270 and the layer normalization 272. Multi-head self attention 250 analogous to the multi-head self attention 206 of FIG. 10A may be applied to the context vector obtained from the layer normalization 272 and the hidden states of the one or more encoder blocks 286, 264. Layer normalization 248 may be applied to the context vector obtained from the multi-head self attention 250 and the context vector obtained from the layer normalization 272 provided via a residual connection. The context vector resulting from the layer normalization 248 may be processed via the feed forward layer 238 and the layer normalization 240 analogous to the description of FIG. 10B. The context vector resulting from the layer normalization 240 may be provided to further decoder blocks 242 analogous to the decoder block 290. The context vector obtained from the one or more decoder blocks 290, 242 may be provided to the decoder output 292. The decoder output 292 may correspond to the decoder output 282 of FIG. 10B.

[0171] With the above-described architecture, the transformer encoder-decoder may receive and process input data at the encoder input 288 and the one or more encoder blocks 286, 264 and the decoder block 290 and the decoder output 292. Based on the input data, the transformer encoder-decoder may generate output data part by part or sequentially. The sequentially generated output data may be provided to and / or may be processed by the decoder input 294, the one or more decoder blocks 290, 242 and the decoder output 292. Preferably, a sequence may be provided to the encoder input 288 and after having generated at least a part of the output data, the decoder input 294 may be provided with at least the part of the elements of the output data already generated. By doing so, the next elements of the output data may be generated with a higher accuracy by taking the input data and the generated output data into account since more data is received by the transformer encoder-decoder may be received over time.

[0172] Because of the transformer encoder-decoder architecture, the transformer encoder-decoder may be configured for transforming a sequence into another representation of the sequence.

[0173] An example for transforming one sequence into another representation may be translation of one sentence into another language. A plurality of transformer encoder-decoders are available in the art such as BART, T5 or the like.

[0174] In an embodiment, the layer normalization 208, 212 may be applied prior to the masked multi-head self attention 224, multi-head self attention 206 and / or the feed forward layer 210 in the transformer decoder, the transformer encoder and / or the transformer encoder-decoder. By doing so, the computational resources for applying the multihead self attention 206 and / or the feed forward layer 210 to the embedded input data and / or the context tensor may be decreased as the entries of the respective tensors may be lower after normalization.

[0175] In an embodiment, the decoder output 292 may comprise of a classification neural network, further feedforward layers, convolutional layers, fully connected layers or the like. For example, the transformer encoder-decoder may be configured for choosing between a plurality of options. For this purpose, the transformer encoder-decoder may be provided with three different input data sets and may classify the context vectors obtained from the one or more decoder blocks 290 via one or more linear layers. Followingly, the architecture may be extended depending on the use case to be solved. [1] FIG. 1 1 illustrates an embodiment of training and / or deploying the transformer encoder, the transformer decoder and / or the transformer encoder-decoder.

[0176] The encoder / decoder / encoder-decoder architecture 302 may correspond to the transformer decoder, the transformer encoder and / or the transformer encoder-decoder as describe within the context of FIG. 10A- FIG. 10C.

[0177] The output data generated by the encoder / decoder / encoder-decoder architecture 302 may comprise of one or more elements, in particular a sequence of elements. The previously generated elements of the output data may be provided as input for generating the next element in the sequence of the output data.

[0178] In the example of FIG. 1 1 , the input data may comprise of N elements, in particular input tokens. An input token may be a token dedicated to be inputted into a data-driven model such as the transformer decoder, the transformer encoder or the transformer encoder-decoder. The output data to be generated may comprise of M elements. The encoder / decoder / encoder-decoder architecture 302 may generate one element of the output data based on receiving the input data and optionally previously generated elements of the output data at a timestep.

[0179] Hence, for generating M elements M time steps are required. A time step comprises of providing input 310, 312, 314 to the encoder / decoder / encoder-decoder architecture 302 and receiving output data 304, 308, 306 from the encoder / decoder / encoder-decoder architecture 302. In a first timestep, the input 310 may comprise of N input tokens. The N input tokens may be associated e.g. with N words, stems or endings. Preferably, the N input tokens may specify a question. One or more input tokens may specify the beginning of the sequence of tokens and / or the end of the sequence of tokens. The input 310 may be processed by the encoder / decoder / encoder-decoder architecture 302. Based on the input 310 at least a part of the output data 304 may be generated. The at least a part of the output data may comprise a first output token. In the next timestep, the generated first output token may be provided together with the input 312. Specifically, where the input 312 may be received by a transformer encoder-decoder the input tokens may be received at the encoder input 288 and the first output token may be received at the decoder input 294. Where the input 312 may be received by the transformer encoder, the input 312 may be received by the encoder input 278 and analogously regarding the transformer decoder and the decoder input 284. Based on the input 312, the output data 308 comprising the first output token and a second output token may be generated. Generating the output data 308 based on the input 312 may refer to generating the second token based on the first token and the N input tokens, wherein the first token may have been generated based on the N input tokens. This process may be repeated until the last token in the sequence of the output data 306 may be generated. Preferably, the last token may be an end token. The end token may terminate the generation of a further output token. Similarly, to the data processing during deployment of the encoder / decoder / encoder-decoder architecture 302, the encoder / decoder / encoder-decoder architecture 302 may be trained. The training data set may comprise a plurality of sequences comprising a plurality of elements. The sequences may be associated with the input data and / or the output data. Additionally or alternatively, the sequences may be independent of the input data and / or the output data. For example, where the input data and the output data may refer to chemical compositions represented via text, the training data set may comprise sequential text data independent of chemical compositions. In this example, the training data set may comprise sequences of words originating from a conversation. In an embodiment, the training data set may comprise at least partially input data sets and / or output data sets.

[0180] The training may be initialized by initializing the encoder / decoder / encoder-decoder architecture 302. In an embodiment, the parameters associated with the encoder / decoder / encoder-decoder architecture 302 may be initialized randomly. Additionally or alternatively, the input embedding of the encoder / decoder / encoder-decoder architecture 302 may be obtained by training a CBOW model or a skip gram model as described within the context of FIG. 9. The trained embedding layer may be used during training. The parameters associated with the embedding layer may be kept constant and / or may be updated after a predefined number of training epochs. By doing so, the number of parameters to be updated is lower enabling a faster and less computational resources- consuming training. Further, the accuracy associated with the embedding layer may be constant and / or may be increased by avoiding error compensation in relation to the just initialized encoder / decoder / encoder-decoder architecture 302.

[0181] During the training of the encoder / decoder / encoder-decoder architecture 302, at least a part of the sequences of the training data set may be provided to the encoder / decoder / encoder-decoder architecture 302 one by another and one or more elements may be generated based on the sequences of the training data set one by another. The elements generated based on the sequences may follow the elements of the parts of sequences the encoder / decoder / encoder-decoder architecture 302 may have been provided with. The generated one or more elements may be compared to the one or more elements following the at least a part of the sequences provided to the encoder / decoder / encoder-decoder architecture 302 as specified by the training data set. Hence, during the training the encoder / decoder / encoder-decoder architecture 302 may generate a guess on the next element and the guess on the next element in a sequence may be compared to the ground truth specifying the actual next element according to the training data set. Based on the guess on the next element and the ground truth a loss may be determined. The loss may define the similarity between the guess on the next element and the ground truth. The loss may be determined by forming a vector dot product between the token associated with the one or more elements and the token associated with the ground truth. A loss unequal to zero may result in updating the parameters associated with encoder / decoder / encoder-decoder architecture 302. Preferably the parameters associated with the encoder / decoder / encoder-decoder architecture 302 may be independent of the embedding layer. For example, the parameters associated with the encoder / decoder / encoder-decoder architecture 302 may be weights of the neurons of the encoder / decoder / encoder-decoder architecture 302.

[0182] Based on the determined loss, backpropagation may be applied to determine the gradients associated with the parameters of the parameters associated with encoder / decoder / encoder-decoder architecture 302 to lower the loss. According to the determined gradients, the parameters associated with the encoder / decoder / encoder- decoder architecture 302, preferably the weights of the neurons associated with the encoder / decoder / encoder- decoder architecture 302, may be updated by using a gradient descent algorithm.

[0183] The training data set may be unlabeled. The sequences of elements within the training data set may inherently comprise the ground truth for determining the loss with respect to the one or more elements generated during the training of the encoder / decoder / encoder-decoder architecture 302. Hence, the encoder / decoder / encoder-decoder architecture 302 may be trained self-supervised. This is advantageous since time and resources for creating a labeled training data set may be saved. Furthermore, this enables the usage of large training data sets associated with a size of several tera bytes. Consequently, the data-driven model may be accurate in generating elements of a sequence. In addition, the large training data set enables few shot predictions or even zero shot predictions. Hence, the data-driven models trained as described above are versatile contributing to saving resources needed for training and / or hosting a plurality of purpose-driven models such as CNNs. The training described above may be referred to as pretraining. The data-driven model may be configured for performing few shot or even zero shot predictions with respect to a plurality of use cases after pretraining. The performance of the data-driven model may be increased further by additional training referred to as finetuning.

[0184] FIG. 12 illustrates an embodiment of input embedding. Where the sequence of elements associated with the input data, preferably comprised in the input data, may be of one type, the input embedding 202, 220, 252, 266 as described within the context of FIG. 10A - FIG. 10C may be used. For example, a type of input data may be text where the elements may be associated with at least a part of a word, a punctuation character, a start token specifying the beginning of one or more sequences associated with the input data and / or the end token. In another example, the input data may be at least partially numerical. Hence, the input data may comprise a plurality of numbers. Numerical input data may be for example tabular data.

[0185] Tabular data may specify one or more rows and / or one or more columns. Hence, the tabular data may comprise one or more cells, wherein the cells may be associated with one or more numerical values.

[0186] Numerical input data may require a different embedding than text input data. Input embeddings for numerical input data may comprise a token embedding, a positional embedding, a column embedding, a row embedding or a combination thereof. Applying a token embedding to one or more elements, in particular tokens associated with the input data may result in a machine-processable representation associated with the one or more elements, in particular tokens. Applying the token embedding to one or more elements may refer to passing the one or more elements through the embedding layer, e.g. as described within the context of FIG. 9. Hence, token embeddings may specify the one or more elements, in particular tokens in a machine-processable representation. For example, the token embedding may transform a numerical value into a vector. This is advantageous since this representation can be enriched by further information such as the position of the token within the sequence and / or within a table associated with the sequence of tokens. The positional embedding may be analogous to the positional embedding as described within the context of FIG. 9, FIG. 10A-FIG. 10C. Where the input data may be tabular data, column embedding may be applied. Applying a column embedding to one or more elements, in particular tokens associated with the input data may result in a machine-processable representation specifying the location of the one or more elements within a table 402, preferably within the columns of the table 402. Applying the column embedding may refer to adding a column factor to the input data embedded via token embeddings, in particular the embedded input data. The column factor may be the same for elements associated with the same column and / or may differ between two or more elements associated with different columns. Analogous, row embeddings may be applied where the input data may be tabular data. Applying a row embedding to one or more elements, in particular tokens associated with the input data may result in a machine-processable representation specifying the location of the one or more elements within a table 402, preferably within the rows of the table 402. Applying the row embedding may refer to adding a column factor to the input data embedded via token embeddings, in particular the embedded input data. The row factor may be the same for elements associated with the same row and / or may differ between two or more elements associated with different rows.

[0187] In an embodiment, input data may be at least partially numerical and at least partially text.

[0188] Hence, the input data may comprise two or more types of data. A type of data may refer to a modality. Followingly, different embeddings may be applied to the input data. To parts of the input data comprising text the input embedding referred to in FIG. 9, FIG. 10A- FIG. 10C may be applied. To parts of the input data being numerical token embeddings, positional embeddings, column embeddings and row embeddings may be applied. Further, segment embeddings may be applied to the input data independent of the type of input data. The segment embedding may specify the type of input data one or more elements may be associated to. For example, if the input data comprises of text and numbers, the input data may comprise of two types of input data. Applying the segment embedding to the input data may refer to adding a segment factor to the input data, preferably the embedded input data and / or the input data after having applied the token embedding. The segment factor may specify the type of data associated with the one or more elements. The segment factor may be the same for one or more elements associated with the same type of input data and / or may differ between two or more elements associated with different types of input data.

[0189] Applying the token embedding, the positional embedding, the segment embedding, the column embedding, the row embedding or a combination thereof may result in embedded input data and / or may be the output of any one of the encoder input 278, 284, 288 or decoder input 284, 294. The data obtained by applying the token embedding, the positional embedding, the segment embedding, the column embedding, the row embedding or a combination thereof may be processed by the encoder block 274, 286, decoder block 280, 290, encoder output 276, decoder output 292, 282.

[0190] The agronomical training data 1013, untrained / pre-trained transformer-based model as described in the context of FIGs. 9-12 and / or the trained transformer-based model as described in the context of FIGs. 2a, 2c, 6 and 8, may be stored in a database, on an electronic data carrier, on a server or in a cloud. An access to the agronomical training data 1013 suitable for training a pre-trained transformer-based model, an untrained / pre-trained transformer-based model as described in the context of FIGs. 9-12 and / or the trained transformer-based model as described in the context of FIGs. 2a, 2c, 6 and 8, may be granted to a user as a computer readable token. The computer readable token may be an authorization key generated by a computer processor upon a request of a requesting computing node associated with a user. The request may be sent to an authorization engine comprising at least one computer processor having rights to grant one or more authorization keys to access the agronomical training data and / or one or more of the untrained, the pre-trained, the trained, the re-trained transformer based models as described withing the context of FIGs. 2a, 2c, 6, 8 and 9-12.

[0191] A user may access the training data via the token for processing the data and receiving the processed result without receiving the actual training data. The user may also receive the actual agronomical training data or a part of the agronomical training data. Depending on the access rights, a user may use (upon receiving the token) one or more of the untrained, pre-trained and / or trained transformer-based models for user’s technical purpose. The user may train / re-train one or more models that the user is granted a permission using the agronomical training data as claimed or using their own training data. An access token for using agronomical training data as claimed may be the same or different as to an access token for using one or more models as described withing the context of FIGs. 2a, 2c, 6, 8 and 9-12.

[0192] Having a separate access token for each of the data products (i.e. , agronomical training data) or data services (i.e., using one or more models as described withing the context of FIGs. 2a, 2c, 6, 8 and 9-12) may increase security aspects of using data products and / or data services. An access token may comprise user credentials, one or more generation algorithms, user authentication, two- factor authentication and / or alike.

[0193] It may be understood that the features of the above embodiments are combinable unless otherwise disclosed in the current application.

[0194] REFERENCE LIST

[0195] 1001 Agronomical resource(s);

[0196] 1002 Equipment;

[0197] 1003 Sensors;

[0198] 1004 Analytics engine;

[0199] 1005 Agronomical data;

[0200] 1006 Control and / or monitoring engine;

[0201] 1007 Machine readable instructions

[0202] 1008 reviewed machine readable instructions;

[0203] 1010 historic agronomic data;

[0204] 1011 Input data;

[0205] 1012 User data;

[0206] 1013 Training data;

[0207] 7001 Prompting Al engine by providing a prompt such as a text prompt (e.g., answer to an inquiry);

[0208] 7002 Adding a context to the prompt (e.g. query data)

[0209] 7003 Optional: adding another context (e.g., a data point from pre-processing engine such as an identification of an agronomical task in metadata, typical action to increase the performance of a field);

[0210] 7004 Generating by Al instructions for monitoring and / or controlling the agronomical resource based on the prompt (e.g., instructions for how to solve and / or answer an agronomical query).

Claims

CLAIMS1 . A computer implemented method for releasing a trained transformer-based model for controlling and / or monitoring an agronomical resource (1001), the method comprising: providing, via a computer interface, agronomical training data associated with one or more agronomical tasks; providing, via the computer interface, a pre-trained transformer-based model comprising at least a transformer component; prompting the pre-trained transformer-based model to re-train the transformer-based model using the agronomical training data associated with the agronomical resource (1001); releasing the trained transformer-based model for controlling and / or monitoring the agronomical resource (1001 ).

2. The computer implemented method of claim 1 , further comprising receiving raw agronomical data (1005) associated with the agronomical resource (1001); pre-processing, by a pre-processing engine, the raw agronomical data associated with the agronomical resource (1001), for providing the agronomical training data associated with the agronomical resource (1001).

3. A computer implemented method for using a trained transformer-based model for controlling and / or monitoring an agronomical resource, the method comprising: releasing a trained transformer-based model for controlling and / or monitoring an agronomical resource (1001 ) according to the method of claims 1 or 2; receiving, by a computer processor, access to the trained transformer based model; receiving, via a computer interface, agronomical input data associated with the one or more agronomical tasks; prompting the trained transformer-based model to analyze the agronomical input data and provide operating instructions (1007) for controlling and / or monitoring the agronomical resource.

4. The computer implemented method of claim 3, further comprising: receiving raw agronomical data associated with the agronomical resource (1001); pre-processing the raw agronomical data by a pre-processing engine (1009) for providing the agronomical input data (101 1).

5. The computer implemented method of claim 3 or 4, wherein the agronomical input data comprise data from a user provided via an interface.

6. The computer implemented method of one of claims 3 to 5, further comprising a prompt from a user for prompting the trained transformer-based model to provide machine readable instructions (1007, 1008) based on a query provided to the model by the user via an interface.

7. The computer implemented method of claim 6, wherein the prompt comprises one or more contexts related to one or more agronomical task at the agronomical environment.

8. The computer implemented method of one of claims 1 -7, wherein the agronomical training data are agronomical data in one or more languages and / or the operating instructions (1007) is in one or more languages.

9. The computer implemented method of one of claims 1 -8, wherein the method further comprises updating the agronomical training data; and prompting the trained transformer-based model to re-train based on the updated agronomical training data.

10. The computer implemented method of claims 9, wherein re-training of the trained transformer-based model is any one of a scheduled re-training, continuous re-training or trigger based-retraining.1 1 . The computer implemented method of claim 10, wherein a trigger for the trigger based-retraining is based on a threshold and / or on a score related to the operating instructions provided by the trained transformerbased model, and / or optionally, a number of re-training cycles.

12. A computer implemented method of any one of claims 3-11 , wherein the operating instructions comprise machine readable instructions (1007, 1008) for controlling and / or monitoring the agronomical resource.

13. A computer program product comprising computer readable instructions that when executed on a computer cause the computer to execute the steps of any one of claims 1 -12.

14. A computer product comprising a computer readable token for accessing agronomical training data and / or the trained or the pre-trained model of any one of claims 1 -12.

15. A cloud computing apparatus comprising a processor for loading computer readable instructions that when executed on the processor cause the processor to execute the method of any one of claims 1 -12.

Images