Reproducibility

All experiments are reproducible via uv.lock + DVC + src/repro.py + environment manifest.

# Quick Start

git clone https://github.com/ltphongssvn/cs1090b_HallucinationLegalRAGChatbots

cd cs1090b_HallucinationLegalRAGChatbots

uv run pre-commit install && uv run pre-commit install --hook-type pre-push

bash setup.sh

uv run python -c "import torch; print(torch.__version__, torch.cuda.is_available())"

uv run dvc pull

uv run pytest --cov=src --cov-report=term-missing

uv.lock (sha256: 3ac3a454...)

Pinned dependency snapshot — 313 packages audited

DVC + S3 (191.64 GB)

34 DVC artifacts, 206 S3 objects, sync clean

src/repro.py

PYTHONHASHSEED=0, CUBLAS=:4096:8, deterministic=True, seed=0 across 4 GPUs

Reproducibility Verification Summary

Cell 1 preflight: 21/21 assertions passed (5 TDD contract + 16 preflight). uv.lock sha256: 3ac3a4547106ba177af6c78046b120bfea376d7d40e8d77d75bede774d0660ee. W&B: 45 offline runs queued (wandb/offline-run-*), lineage DAG complete. Final artifact SHA-256: 43eec4d3023f9485... DVC: 34 artifacts, 191.64GB, 206 S3 objects, sync clean. git_sha: b16f03f76a08.