Non-instructional Fine-tuning: Enabling Instruction-Following Capabilities in Pre-trained Language Models without Instruction-Following Data
Fuente:
arXiv
Enregistré dans:
| Auteurs principaux: | , , |
|---|---|
| Format: | Preprint |
| Publié: |
2024
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866912009909436416 |
|---|---|
| author | Xie, Juncheng Syu, Shensian Lee, Hung-yi |
| author_facet | Xie, Juncheng Syu, Shensian Lee, Hung-yi |
| contents | Instruction fine-tuning is crucial for today's large language models (LLMs) to learn to follow instructions and align with human preferences. Conventionally, supervised data, including the instruction and the correct response, is required for instruction fine-tuning. To obtain such data, some researchers prompted well-trained models like GPT-4 to generate instructions and correct responses. In this paper, we propose a novel approach that uses the first half of a random text from OpenWebText as the instruction and GPT-3.5-turbo or GPT-4-turbo to complete the text as the response. Despite the data being "non-instructional", we found that pre-trained LLMs fine-tuned on this data can gain instruction-following capabilities. This observation is verified by fine-tuning several well-known pre-trained LLMs (e.g., LLaMA-2-7B, LLaMA-3-8B, LLaMA-3-70B, Mistral-7B-v0.1). The "non-instructional data" also improved some models that underwent supervised fine-tuning and human preference alignment. Our LLaMA-3-70B-Instruct fine-tuned through "non-instructional data" is comparable with LLaMA-3.1-70B-Instruct on the Arena Hard leaderboard. We analyzed the "non-instructional data" and ensured it is devoid of content related to instruction fine-tuning. Our findings will inspire further investigation into how to develop instruction-following capabilities without explicit instruction-related data. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2409_00096 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Non-instructional Fine-tuning: Enabling Instruction-Following Capabilities in Pre-trained Language Models without Instruction-Following Data Xie, Juncheng Syu, Shensian Lee, Hung-yi Computation and Language Artificial Intelligence Machine Learning Instruction fine-tuning is crucial for today's large language models (LLMs) to learn to follow instructions and align with human preferences. Conventionally, supervised data, including the instruction and the correct response, is required for instruction fine-tuning. To obtain such data, some researchers prompted well-trained models like GPT-4 to generate instructions and correct responses. In this paper, we propose a novel approach that uses the first half of a random text from OpenWebText as the instruction and GPT-3.5-turbo or GPT-4-turbo to complete the text as the response. Despite the data being "non-instructional", we found that pre-trained LLMs fine-tuned on this data can gain instruction-following capabilities. This observation is verified by fine-tuning several well-known pre-trained LLMs (e.g., LLaMA-2-7B, LLaMA-3-8B, LLaMA-3-70B, Mistral-7B-v0.1). The "non-instructional data" also improved some models that underwent supervised fine-tuning and human preference alignment. Our LLaMA-3-70B-Instruct fine-tuned through "non-instructional data" is comparable with LLaMA-3.1-70B-Instruct on the Arena Hard leaderboard. We analyzed the "non-instructional data" and ensured it is devoid of content related to instruction fine-tuning. Our findings will inspire further investigation into how to develop instruction-following capabilities without explicit instruction-related data. |
| title | Non-instructional Fine-tuning: Enabling Instruction-Following Capabilities in Pre-trained Language Models without Instruction-Following Data |
| topic | Computation and Language Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2409.00096 |