3 results about "False rejection" patented technology

This invention discloses a method and system for suppressing false rejection behavior and selectively forgetting harmful semantics in large language models across vertical domains based on feature subspace decoupling, belonging to the field of fine-tuning and alignment technology for large artificial intelligence models. The method includes steps for target definition and data construction, feature subspace localization, key level optimization, orthogonal decoupling weight correction, constrained norm renormalization, selective fine-tuning, false rejection calibration, and closed-loop verification monitoring and document output. The system includes modules for concept discovery and data construction, representation subspace modeling, false rejection localization and calibration, and verifiable evaluation and continuous monitoring. This invention can significantly reduce the false rejection rate and maintain a safety baseline while preserving the model's generalizability, meeting the compliance requirements of generative artificial intelligence services.

This invention provides a method for constructing high-quality metal material process datasets based on a large language model, relating to the field of process data collection and processing technology. It introduces a chemical composition hashing (HC) mechanism, which transforms complex composition information into unique identifiers through standardized concatenation of a preset element set and MD5 hashing, enabling efficient and accurate cross-file comparison. A three-level duplicate detection system—precise duplicates → similar duplicates → LLM intelligent confirmation—effectively avoids false rejections due to differences in expression while ensuring high recall. A prompting engineering system is constructed based on a locally deployed large language model, integrating domain knowledge templates and strict output format constraints. This achieves complete reconstruction and structured expression of complex process chains without manual intervention, enabling end-to-end extraction of standardized material entries from massive amounts of unstructured documents. Furthermore, semantic-level feature hashing and LLM-assisted judgment mechanisms significantly improve deduplication accuracy and logical consistency.

The application relates to the technical field of network security and artificial intelligence, and specifically discloses a damage vulnerability sample generation method based on self-consistent explanation, which first extracts damage scene features from multi-source situation data, maps the features into a structured thinking chain of syllogistic logical reasoning of a forced large model according to vulnerability principle-triggering condition-damage consequence, generates codes after establishing a complete cause-effect chain from the source constraint model, and avoids statistical probability-driven shallow imitation. Subsequently, the generated code segments and mechanism assertion texts are respectively subjected to static topology reverse deduction and semantic coding, cross self-consistency verification is realized by calculating cosine similarity, and false defect samples are filtered. Further, an attack chain cascading dependence graph and link propagation enhancement scoring are introduced, the cause-effect conduction relationship among samples is brought into evaluation, false rejection of attack chain bridging nodes and false retention of logically contradictory samples are avoided, and finally a high-fidelity vulnerability sample library that can be used for multi-level damage scene testing in an industrial field is generated.