AI Training Data Filtering for Hallucination Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Generative AI models can produce low-quality or incorrect answers due to overfitting from training data with identical or redundant text sequences, leading to 'hallucinations' and decreased model reliability.
Innovation Solution
Implement a procedure to analyze and filter training data records for artificial intelligence models, discarding or assigning lower weights to data records generated by AI or containing overlaps, thereby improving the quality and reliability of AI outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training data contains multiple identical or similar text sequences, then the model learns patterns more effectively, but the model assigns too high probability to these sequences leading to overfitting and hallucinations
Solution Approach 1:
The patent applies preliminary action by detecting and removing duplicate or similar training data records before the model training process begins. The system identifies redundant text sequences in the training corpus and eliminates them in advance, preventing the model from overfitting to repeated patterns. This pre-processing step ensures that the model learns from diverse, high-quality data without being biased by redundant information.
2Quantity of substance
If the model is trained on all available training data, then more information is learned, but the model produces lower quality answers due to poor-quality or redundant data
Solution Approach 1:
The patent applies local quality by implementing selective data filtering that treats different training data records differently based on their quality and uniqueness. Rather than uniformly processing all data, the system identifies and removes specific low-quality or redundant records while preserving high-quality diverse data. This selective approach ensures that the model trains on the most valuable information, improving answer quality without sacrificing the benefits of large-scale training.
3Adaptability or versatility
If the model generates answers based on probability distributions from training data, then it can respond to any query, but it may generate incorrect answers when training data contains errors or biases
Solution Approach 1:
The patent applies preliminary action by cleaning and de-duplicating the training corpus before model training. By removing erroneous, biased, or redundant data records in advance, the system prevents these quality issues from being encoded into the model's probability distributions. This ensures that the model maintains its ability to respond to diverse queries while generating more accurate and reliable answers based on cleaned training data.
Data Source
Figure 1
AI summary
The present invention relates to techniques for improving models of an artificial intelligence, in particular a generative AI, comprising the following steps: • Maintaining a model f of an artificial intelligence, wherein the model f is to be trained, in particular to generate a new version; • Maintaining datasets for training the model f; • Analyzing the datasets with regard to their suitability for training by an algorithm or the artificial intelligence; • Discarding or lowering the weight of a dataset if the algorithm or the artificial intelligence determines that the dataset was generated by the artificial intelligence or another artificial intelligence, or that the dataset has overlaps with other datasets.