.gitignore

# =============================================================================
# .gitignore — what NOT to commit
# -----------------------------------------------------------------------------
# Each block below is grouped by *why* the files are excluded so future readers
# (and recruiters) understand the engineering rationale, not just the patterns.
# =============================================================================

# ---- Python bytecode / packaging ---------------------------------------------
# Compiled artefacts. Regenerated automatically on every run.
__pycache__/
*.py[cod]
*.egg-info/
*.egg
.eggs/
build/
dist/
pip-wheel-metadata/

# ---- Virtual environments ----------------------------------------------------
# Per-developer; pinning is done via requirements.txt + .python-version.
venv/
.venv/
env/
.env-tf/

# ---- Python tooling caches ---------------------------------------------------
# Speed up local runs; nothing portable. Caches are recreated by the tools.
.pytest_cache/
.mypy_cache/
.ruff_cache/
.tox/
.coverage
.coverage.*
htmlcov/
coverage.xml
.nox/
.hypothesis/

# ---- Jupyter / notebooks -----------------------------------------------------
# Checkpoints are autosaves; outputs are stripped by `nbstripout` pre-commit
# so notebook diffs stay reviewable.
.ipynb_checkpoints/
*.ipynb_checkpoints

# ---- ML / experiment tracking ------------------------------------------------
# MLflow's local store, model artefacts, training run dumps. These are large
# and should live in a model registry (HuggingFace Hub) or experiment-tracking
# server (DagsHub MLflow), not in Git.
mlruns/
mlartifacts/
outputs/
runs/
wandb/
lightning_logs/

# ---- Model weights / serialised artefacts ------------------------------------
# Large binaries — published via HuggingFace Hub, not Git.
*.h5
*.keras
*.pt
*.pth
*.ckpt
*.onnx
*.tflite
*.pb
*.savedmodel/
*.safetensors

# ---- Tokenizer / vocabulary artefacts ----------------------------------------
# Pickles can carry RCE risk if blindly loaded from untrusted sources.
# Vocabularies are versioned alongside their model in models/<version>/.
*.pkl
vocab_*.file
sentencepiece.model

# ---- Datasets ----------------------------------------------------------------
# COCO is downloaded by `scripts/prepare_data.py`; never committed.
data/
datasets/
*.tfrecord
*.tfrecords
# Source package `src/captioning/data/` is library code, not a dataset folder.
!src/captioning/data/
!src/captioning/data/*.py

# ---- Environment / secrets ---------------------------------------------------
# `.env.example` is committed as the schema; `.env` never is.
.env
.env.local
.env.*.local
!.env.example

# ---- Node / frontend ---------------------------------------------------------
node_modules/
.next/
.turbo/
.vercel/
out/
*.tsbuildinfo
npm-debug.log*
yarn-debug.log*
yarn-error.log*

# ---- Docker / build ----------------------------------------------------------
.docker/

# ---- Editors / IDEs ----------------------------------------------------------
# Per-developer settings. Workspace-shared settings should go in `.vscode/*.json`
# explicitly committed; anything else stays local.
.vscode/
.idea/
*.swp
*.swo
*~

# ---- OS noise ----------------------------------------------------------------
.DS_Store
Thumbs.db
desktop.ini

# ---- Claude / AI tooling -----------------------------------------------------
# Local Claude Code session state. Contains user-specific settings.
.claude/
frontend/.vite/