.gitignore

# One .gitignore for the whole repo — every pipeline folder is a separate uv
# project, but none of them need their own ignore file; patterns below are
# unanchored so they match at any depth (pre-training/.venv/, fine-tuning/
# llava15-lora/.venv/, etc. are all covered by a single `.venv/` line).

# --- Python / uv ---
.venv/
__pycache__/
*.pyc
*.egg-info/

# uv.lock IS committed for reproducibility — do not ignore it.
# .python-version IS committed — it pins the interpreter uv uses per project.

# --- All downloads / caches live under .cache/ (HF_HOME, UV_CACHE_DIR) ---
# so cleanup is just `rm -rf .cache/` — nothing else on disk gets touched.
.cache/

# --- Hugging Face cache for base models (redownloaded on first run) ---
hf_cache/

# --- Training/eval runtime noise and large binaries: never committed ---
wandb/
*.safetensors
*.bin
*.pt
*.ckpt
last_run_prepared/

# --- Fused/merged models (regenerable via each pipeline's merge script) ---
merged_model/

# --- Raw/prepared datasets: drop zones, scripts fetch/generate them locally.
# Each of these keeps a committed README.md as the only tracked file.
# Patterns with an internal slash anchor to the .gitignore's own directory
# (repo root here), so they need an explicit **/ prefix to also match inside
# fine-tuning/<pipeline>/, serving/<pipeline>/, etc. ---
**/DATASET/*
!**/DATASET/README.md
**/DATASET_JSONL/*
!**/DATASET_JSONL/README.md
**/data/*
!**/data/README.md
**/runs/*
!**/runs/README.md
**/output/*
!**/output/README.md
**/outputs/*
!**/outputs/README.md

# --- pre-training: PDF corpus + rendered pages + OCR CSVs ---
*.png
*.pdf
*.csv
Release_1/
Release_1_PNG/
conversion_log.txt