Training Data Identification via N-gram Source Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional language learning models (LLMs) do not reveal the training data used for their development, making it difficult to identify potential issues such as biases or intellectual property infringement.
Innovation Solution
A method that queries a language learning model, generates groupings of the output response into n-grams, and calculates a source score based on two different training datasets to identify the training data used by the model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional LLMs are designed to understand and reproduce human language by analyzing training data, then the model's language processing capability is improved, but the transparency of training data is lost
Solution Approach 1:
The patent introduces an intermediary analysis system that acts as a mediator between the LLM and the training data. This system queries the LLM's responses, extracts n-grams, compares them against candidate training datasets, and identifies potential training data sources without requiring the LLM itself to disclose its training data. The intermediary system bridges the gap between maintaining model capabilities and achieving training data transparency.
2Ease of operation
If the LLM does not reveal the training data, then the model's operational simplicity is maintained, but the ability to detect biases and IP infringement is reduced
Solution Approach 1:
The patent segments the training data into discrete candidate datasets and systematically compares n-grams from model responses against each segment. By breaking down the training data identification process into manageable comparisons between response n-grams and candidate dataset n-grams, the system can detect biases and IP infringement without requiring the entire LLM to be transparent about its training data, thus maintaining operational simplicity while enabling harmful factor detection.
3Loss of information
If training data identification methods are implemented, then the transparency of training data is improved, but the system complexity increases
Solution Approach 1:
The patent implements a self-service approach where the LLM itself is used to generate the responses that will be analyzed to identify training data. The model queries itself with questions about its training data, and its own responses are then processed to extract n-grams for comparison. This self-service mechanism reduces the need for external complex analysis systems while still achieving training data transparency.
Data Source
AI summary
Techniques are provided for training data identification and model selection. In one embodiment, the techniques involve querying a first language learning model with a first query, wherein the first language learning model generates a text output response to the first query, generating groupings of the text output response, wherein the groupings include multiple n-grams, generating a source score of the groupings based on a first training data and a second training data, and identifying the first training data as training data of the first language learning model based on the source score.