# Core
torch>=2.0.0
transformers>=4.35.0
datasets>=2.14.0
accelerate>=0.24.0

# Metrics
scikit-learn>=1.3.0
scipy>=1.11.0

# Optional but recommended
trackio>=0.1.0

# FlashAttention (requires CUDA)
# flash-attn>=2.3.0

# For longer context / efficiency
# xformers>=0.0.22