__pycache__/
*.py[cod]
checkpoints/
data/
output*/
*.log
wandb/

# Large/generated artifacts
*.pt
*.zip
*.png
!docs/img/*.png
blog_review_first100/
blog_review_first100_light/
analysis/
tmp_*.json
hf_dataset_export/
hf_dataset_export_sharded/

# One-off data prep/upload scripts (data already on HF)
prepare_andy_datasets.py
upload_andy_datasets.py
package_andy_shards.py

# Training outputs / checkpoint symlinks
training/output*
training/eval_results

# Embedding caches (large, regenerable)
tmp/*.pkl

# Local scratch / historical artifacts (eval JSONLs land in --output-dir)
tmp/
eval_output/

# Serving venvs
serving/vllm/.venv/
.env

# Eval tiles / local caches
test_miniv6/
test_miniv7/
test_miniv8/

# HF upload staging (already on Hub)
hf_upload_*/

# Claude Code local state
.claude/

# SFT: LlamaFactory fork venv + outputs (11 GB)
sft/LlamaFactory/

# Keep eval JSONs (useful eval artifacts), skip tqdm log noise
sft/eval_out/*.log