Teaching LLMs new tricks without making them forget their old knowledge
NILE, proposed in this paper, framework enhances instruction fine-tuning by aligning LLMs' internal knowledge with training datasets, leading to significantly improved model performance.
-----
https://arxiv.org/abs/2412.16686
🤔 Original Problem:
Existing instruction fine-tuning datasets often contain knowledge inconsistent with LLMs' internal knowledge from pre-training, reducing training effectiveness and limiting model capabilities.
-----
🔧 Solution in this Paper:
→ NILE framework extracts LLMs' internal knowledge corresponding to instruction data using few-shot prompting and demonstration learning
→ Knowledge-aware Sample Revision component revises training samples by infusing extracted internal knowledge
→ Internal Consistency Filtering measures and filters samples based on their alignment with LLM's internal knowledge
→ The framework maintains optimal balance between consistent and inconsistent knowledge in training data
-----
💡 Key Insights:
→ Internal knowledge consistency is crucial for unlocking LLM capabilities
→ Balancing consistent and inconsistent knowledge improves generalization
→ Few-shot demonstration learning effectively extracts internal knowledge
-----
📊 Results:
→ 66.6% performance gain on Arena-Hard benchmark
→ 68.5% improvement on Alpaca-Eval V2
→ Significant boosts in BBH tasks: 4.64 points for Mistral and 1.05 for Llama-3
