Datasets:
| # Script Fidelity Benchmark - Python dependencies | |
| # | |
| # IMPORTANT: datasets must be <=2.x. Version 3+ removed support for dataset | |
| # scripts, which google/fleurs still requires (fleurs.py). Installing 4.x will | |
| # cause a RuntimeError on load_dataset('google/fleurs', ...). | |
| # | |
| # IMPORTANT: do NOT install torch from the default package index. It will pull | |
| # the latest CUDA build (cu130+) which requires driver >=550. Most cloud GPU | |
| # pods (vast.ai, RunPod) ship driver ~520 (CUDA 12.2). Use the cu121 index: | |
| # | |
| # uv pip install torch --index-url https://download.pytorch.org/whl/cu121 | |
| # | |
| # If torch.cuda.is_available() returns False and you see | |
| # "NVIDIA driver too old (found version 12020)", this is the cause. | |
| # | |
| # Tested on Python 3.10, CUDA 12.1, driver 520, Tesla P100 16 GB. | |
| datasets==2.21.0 | |
| fsspec[http]<=2024.6.1 | |
| evaluate>=0.4.0,<1.0 | |
| # transformers 5.x required for AutoModelForMultimodalLM (Gemma 4 audio support) | |
| transformers>=5.0.0 | |
| protobuf>=4.25.0,<6 | |
| sentencepiece>=0.1.99 | |
| # torch: install separately; see note above | |
| torch>=2.5.0 | |
| tqdm>=4.65.0 | |
| numpy>=1.24.0 | |
| pandas>=2.0.0 | |
| matplotlib>=3.7.0 | |
| seaborn>=0.13.0 | |
| huggingface_hub>=0.23.0 | |
| soundfile>=0.12.0 | |
| librosa>=0.10.0 | |
| jiwer>=3.0.0 | |
| sacrebleu>=2.4.0 | |
| langdetect>=1.0.9 | |
| # Required for Gemma 4 multimodal loading | |
| accelerate>=0.27.0 | |
| torchvision>=0.15.0 | |