Work xTuring - LLM Fine-Tuning Library
xTuring - LLM Fine-Tuning Library
2023
Summary
Open-sourced a parameter-efficient LLM fine-tuning library supporting LoRA, QLoRA, and quantized training for major open-source models.
My Role
Project Lead & Engineering Lead
Challenges
- Unified interface across models with different chat templates
- Package installation conflicts across GPU/CPU systems
- Stability issues with 4-bit and 8-bit quantization
What We Did
- Standardized APIs for fine-tuning across LLaMA, Falcon, etc.
- Built dataset preprocessing tools to generate instruction datasets
- Managed pip packaging, community Discord, and documentation
Outcomes
- 2.7k+ GitHub stars
- Used by hundreds of fine-tuning practitioners
- Enabled quick experimentation with minimal infra