# Supplementary materials — analysis-script dependencies
# Match versions used in the paper's evaluation pipeline (see Appendix D)

lm-evaluation-harness==0.4.12
datasets>=3.0,<4.0
transformers>=4.50,<6.0
vllm>=0.10
numpy
matplotlib