Understanding Catastrophic Forgetting in Language Models via Implicit Inference

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Kotha, Suhas, Springer, Jacob Mitchell, Raghunathan, Aditi
Format: Preprint
Published: 2023
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913312783990784
author Kotha, Suhas
Springer, Jacob Mitchell
Raghunathan, Aditi
author_facet Kotha, Suhas
Springer, Jacob Mitchell
Raghunathan, Aditi
contents We lack a systematic understanding of the effects of fine-tuning (via methods such as instruction-tuning or reinforcement learning from human feedback), particularly on tasks outside the narrow fine-tuning distribution. In a simplified scenario, we demonstrate that improving performance on tasks within the fine-tuning data distribution comes at the expense of capabilities on other tasks. We hypothesize that language models implicitly infer the task of the prompt and that fine-tuning skews this inference towards tasks in the fine-tuning distribution. To test this, we propose Conjugate Prompting, which artificially makes the task look farther from the fine-tuning distribution while requiring the same capability, and we find that this recovers some of the pretraining capabilities in our synthetic setup. Since real-world fine-tuning distributions are predominantly English, we apply conjugate prompting to recover pretrained capabilities in LLMs by simply translating the prompts to different languages. This allows us to recover in-context learning abilities lost via instruction tuning, natural reasoning capability lost during code fine-tuning, and, more concerningly, harmful content generation suppressed by safety fine-tuning in chatbots like ChatGPT.
format Preprint
id arxiv_https___arxiv_org_abs_2309_10105
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Understanding Catastrophic Forgetting in Language Models via Implicit Inference
Kotha, Suhas
Springer, Jacob Mitchell
Raghunathan, Aditi
Computation and Language
Machine Learning
We lack a systematic understanding of the effects of fine-tuning (via methods such as instruction-tuning or reinforcement learning from human feedback), particularly on tasks outside the narrow fine-tuning distribution. In a simplified scenario, we demonstrate that improving performance on tasks within the fine-tuning data distribution comes at the expense of capabilities on other tasks. We hypothesize that language models implicitly infer the task of the prompt and that fine-tuning skews this inference towards tasks in the fine-tuning distribution. To test this, we propose Conjugate Prompting, which artificially makes the task look farther from the fine-tuning distribution while requiring the same capability, and we find that this recovers some of the pretraining capabilities in our synthetic setup. Since real-world fine-tuning distributions are predominantly English, we apply conjugate prompting to recover pretrained capabilities in LLMs by simply translating the prompts to different languages. This allows us to recover in-context learning abilities lost via instruction tuning, natural reasoning capability lost during code fine-tuning, and, more concerningly, harmful content generation suppressed by safety fine-tuning in chatbots like ChatGPT.
title Understanding Catastrophic Forgetting in Language Models via Implicit Inference
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2309.10105