Rule-based data generation outperforms LLM-based approaches for function call training
Alopex enables precise on-device function calls while preserving LLM's general abilities
https://arxiv.org/abs/2411.05209
Original Problem 🤔:
Function call capabilities in on-device LLMs face three major challenges: scarce training data requiring manual verification, ineffective question formatting leading to inaccuracies, and catastrophic forgetting of general abilities after fine-tuning.
-----
Solution in this Paper 🛠️:
→ Introduces Alopex framework with a Rule-Based Logic approach for generating high-quality training data without manual verification
→ Implements a novel "description-question-output" format that outperforms existing approaches and reduces function information leakage
→ Uses a 1:1 data mixing strategy with textbook datasets to prevent catastrophic forgetting while maintaining general capabilities
→ Achieves automated adaptation pipeline for data generation and LLM fine-tuning
-----
Key Insights 🔍:
→ Rule-Based Logic generates training data with 99% accuracy compared to 70% for LLM-based generation
→ Placing function descriptions before questions improves out-of-logic accuracy
→ Mixing function call data with textbook datasets in 1:1 ratio preserves general capabilities
→ Works effectively with smaller LLMs (1.6B-2B parameters)
-----
Results 📊:
→ Achieves 99% function call accuracy across multiple models
→ Maintains performance on general tasks (MMLU, GSM8K, etc.)
→ Significantly reduces catastrophic forgetting compared to baseline
→ Fox-1-1.6B shows highest robustness and average accuracy among tested models
Share this post