Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

CoT-Self-Instruct: Building high-quality synthetic prompts for reasoning and non-reasoning tasks

Created by
  • Haebom

Author

Ping Yu, Jack Lanchantin, Tianlu Wang, Weizhe Yuan, Olga Golovneva, Ilia Kulikov, Sainbayar Sukhbaatar, Jason Weston, Jing Xu

Outline

This paper proposes CoT-Self-Instruct, a high-quality synthetic data generation method for large-scale language models (LLMs). Based on a seed task, CoT-Self-Instruct first infers and plans the LLM through Chain-of-Thought (CoT), and then generates new synthetic data of similar quality and complexity. This is followed by a filtering step that selects high-quality data using automatic evaluation metrics, and the selected data is used for LLM training. Experimental results show that CoT-Self-Instruct outperforms existing training datasets (s1k, OpenMathReasoning) on verifiable reasoning tasks (MATH500, AMC23, AIME24, GPQA-Diamond) and human-generated data and standard self-instruct training data on non-verifiable instruction-following tasks (AlpacaEval 2.0, Arena-Hard).

Takeaways, Limitations

•
Takeaways:
◦
CoT-Self-Instruct can contribute to improving the performance of LLM by generating synthetic data with higher quality than existing datasets.
◦
It shows excellent performance on both verifiable inference and unverifiable instruction following tasks.
◦
We present a method for efficiently selecting high-quality data using automated evaluation metrics.
•
Limitations:
◦
Further validation of the generalization performance and reliability of the proposed automatic evaluation metric is needed.
◦
There is a lack of analysis of the bias and safety of the generated synthetic data.
◦
Since performance was evaluated only for a specific type of task, generalization performance to other types of tasks needs to be examined.
👍
Made with Slashpage