Don't Half-listen: Capturing Key-part Information in Continual Instruction Tuning
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2024
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866908380782657536 |
|---|---|
| author | He, Yongquan Zhang, Wenyuan Huang, Xuancheng Zhang, Peng Meng, Lingxun Zhou, Xiang Zeng, Ke Cai, Xunliang |
| author_facet | He, Yongquan Zhang, Wenyuan Huang, Xuancheng Zhang, Peng Meng, Lingxun Zhou, Xiang Zeng, Ke Cai, Xunliang |
| contents | Instruction tuning for large language models (LLMs) can drive them to produce results consistent with human goals in specific downstream tasks. However, the process of continual instruction tuning (CIT) for LLMs may bring about the catastrophic forgetting (CF) problem, where previously learned abilities are degraded. Recent methods try to alleviate the CF problem by modifying models or replaying data, which may only remember the surface-level pattern of instructions and get confused on held-out tasks. In this paper, we propose a novel continual instruction tuning method based on Key-part Information Gain (KPIG). Our method computes the information gain on masked parts to dynamically replay data and refine the training objective, which enables LLMs to capture task-aware information relevant to the correct response and alleviate overfitting to general descriptions in instructions. In addition, we propose two metrics, P-score and V-score, to measure the generalization and instruction-following abilities of LLMs. Experiments demonstrate our method achieves superior performance on both seen and held-out tasks. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2403_10056 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Don't Half-listen: Capturing Key-part Information in Continual Instruction Tuning He, Yongquan Zhang, Wenyuan Huang, Xuancheng Zhang, Peng Meng, Lingxun Zhou, Xiang Zeng, Ke Cai, Xunliang Computation and Language Artificial Intelligence Instruction tuning for large language models (LLMs) can drive them to produce results consistent with human goals in specific downstream tasks. However, the process of continual instruction tuning (CIT) for LLMs may bring about the catastrophic forgetting (CF) problem, where previously learned abilities are degraded. Recent methods try to alleviate the CF problem by modifying models or replaying data, which may only remember the surface-level pattern of instructions and get confused on held-out tasks. In this paper, we propose a novel continual instruction tuning method based on Key-part Information Gain (KPIG). Our method computes the information gain on masked parts to dynamically replay data and refine the training objective, which enables LLMs to capture task-aware information relevant to the correct response and alleviate overfitting to general descriptions in instructions. In addition, we propose two metrics, P-score and V-score, to measure the generalization and instruction-following abilities of LLMs. Experiments demonstrate our method achieves superior performance on both seen and held-out tasks. |
| title | Don't Half-listen: Capturing Key-part Information in Continual Instruction Tuning |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2403.10056 |