Multilingual Dataset Collection For Large Language Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing large language models (LLMs) trained predominantly in English face suboptimal performance in non-English contexts due to imbalanced training datasets and inefficiencies in current multilingual dataset creation methods, such as translating and templating, which introduce errors and lack diversity.
Innovation Solution
A novel framework using English-focused LLMs, monolingual corpora, and a scoring function to generate high-quality, diverse instruction-following task (IFT) datasets by selecting responses in target languages, translating them into English, and evaluating their quality to ensure naturalness and diversity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing LLMs are trained predominantly in English, then English language performance is improved, but non-English language performance deteriorates
Solution Approach 1:
The patent segments the training data collection process into distinct stages: collecting monolingual responses in target languages, translating them to English, generating instructions, scoring pairs, and translating instructions back. This segmentation allows each stage to be optimized independently while maintaining overall multilingual training quality.
Solution Approach 2:
The patent inverts the traditional translation approach by first collecting high-quality responses in the target language, then translating them to English for instruction generation, and finally translating instructions back to the target language. This inversion ensures the target language content remains authentic while benefiting from English-language model capabilities.
2Adaptability or versatility
If translation methods are used to create multilingual datasets, then dataset coverage is improved, but data quality deteriorates due to translation errors
Solution Approach 1:
The patent uses English as an intermediary language in a mediating role. Target language responses are translated to English, processed through the English LLM for instruction generation, then translated back. This intermediary approach allows leveraging English-language model capabilities while preserving target language authenticity through the bidirectional translation process.
Solution Approach 2:
The patent implements a scoring mechanism that evaluates instruction-response pairs and filters out low-quality pairs before final translation. This feedback loop ensures that only high-quality pairs are translated back to the target language, maintaining data quality while achieving broad language coverage.
3Productivity
If templating methods are used to create multilingual datasets, then dataset generation speed is improved, but data diversity deteriorates
Solution Approach 1:
The patent enables the system to automatically collect monolingual responses from various sources, translate them, generate instructions, score pairs, and translate back without manual intervention. This self-service approach maintains high productivity while achieving diversity through automatic collection from multiple response sources rather than relying on fixed templates.
4Manufacturing precision
If more translation steps are added to preserve linguistic nuances, then language authenticity is improved, but processing complexity increases
Solution Approach 1:
The patent segments the complex processing pipeline into distinct, manageable stages: response collection, translation to English, instruction generation, pair scoring, filtering, translation back to target language, and dataset compilation. This segmentation makes the complex process more manageable and allows each stage to be optimized independently while maintaining linguistic authenticity.
Data Source
AI summary
Multilingual dataset collection for large language model (LLM) training is performed to prepare a dataset in each of multiple target languages according to source responses obtained and high-quality instructions generated for those responses. Responses in the target language from one or more response sources and translated into English to produce English responses. For each English response, a first LLM is prompted to generate an English instruction for which the English response is a valid answer. For each English response and corresponding instruction pair, a score for the pair is compared against a threshold, and, where the score meets the threshold, the English instruction of the pair is translated into the target language, in which each instruction in the target language and the respective response in the target language are a final pair. A training dataset including the final pairs is then output for training a second LLM.