| # Core training stack | |
| # Tested API targets: TRL v1.3+, Transformers v5.5+/v5.6+, PEFT v0.19+, Datasets v4.8+ | |
| torch>=2.6 | |
| transformers>=5.5.0 | |
| trl[peft]>=1.3.0 | |
| peft>=0.19.0 | |
| accelerate>=1.12.0 | |
| datasets>=4.8.0 | |
| bitsandbytes>=0.48.0 | |
| safetensors>=0.5.0 | |
| huggingface_hub>=1.0.0 | |
| trackio>=0.8.0 | |
| # Evaluation / utilities | |
| pandas>=2.2.0 | |
| numpy>=2.0.0 | |
| tqdm>=4.66.0 | |
| jsonschema>=4.23.0 | |
| scikit-learn>=1.5.0 | |
| pyyaml>=6.0.2 | |
| rich>=13.7.0 | |
| # Optional but recommended on RTX 6000 Ada for packing/throughput. | |
| # Install separately if your CUDA/PyTorch build supports it: | |
| # pip install flash-attn --no-build-isolation | |