New SFT-as-Context method mitigates LLM forgetting during fine-tuning

Researchers have introduced a novel training-free method called SFT-as-Context to address the issue of catastrophic forgetting in supervised fine-tuning (SFT) of large language models (LLMs). This technique allows a parent model to leverage the SFT model's responses as context, enabling it to acquire specialized capabilities through in-context learning while retaining its original general knowledge. Experiments across numerous model pairs and benchmarks demonstrated that SFT-as-Context maintains performance close to SFT models on fine-tuned tasks and near parent models on general capabilities, effectively bridging the gap for queries requiring both. AI

IMPACT This method could improve the efficiency and effectiveness of fine-tuning LLMs, enabling them to better handle complex queries requiring both specialized and general knowledge.

RANK_REASON The cluster contains an academic paper detailing a new method for fine-tuning LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• AIME 2024
• Hugging Face
• large language models
• LiveCodeBench
• NutriBench-English
• SFT-as-Context
• supervised fine-tuning

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →