.gitignore

venv/
.venv/
__pycache__/
**/__pycache__/
*.pyc
.env
*.env
.env.*
!.env.example
**/.env

models/

# Redirected huggingface_hub cache (learnmate/config.py); avoids ~/.cache/huggingface on
# machines where that directory isn't writable. Downloaded models, not source.
integrated-backend/data/hf_cache/

# Scratch space for scripts/build_finetuned_gguf.py: the LoRA adapter merged into its base
# as a ~3.1 GB fp16 model, plus the fetched llama.cpp converter. Entirely reproducible from
# the adapter and the base weights, and deletable once the GGUF in models/ exists.
integrated-backend/build/

# Runtime evaluation log; regenerated on every run
scores.jsonl

# Qdrant embedded-mode lock file
*.lock

*.zip

# Server and dev-server logs; written on every run, never worth a merge conflict
*.log

# model-Thevindu local artifacts / secrets / corpus
model-Thevindu/01_dataset_pipeline/data/intermediate/
model-Thevindu/01_dataset_pipeline/data/output/
model-Thevindu/01_dataset_pipeline/processed*/
model-Thevindu/01_dataset_pipeline/data/excluded_from_corpus/
model-Thevindu/03_testing_and_versioning/New/
model-Thevindu/01_dataset_pipeline/data/raw_pdfs/*
!model-Thevindu/01_dataset_pipeline/data/raw_pdfs/.gitkeep
!model-Thevindu/01_dataset_pipeline/data/raw_pdfs/README.md


# Trained weights and training state, wherever they land. The previous rule only
# covered adapters/*/, so a download unpacked anywhere else (e.g. Real_Weights/)
# was silently committable -- ~200 MB of binaries per run, permanent in history.
# The auditable record of a run is run_records/<run_id>.json, which stays tracked.
*.safetensors
*.bin
*.pt
*.pth
model-Thevindu/02_finetuning/**/checkpoints/

# macOS Finder metadata. Committed by accident on the Keycloak branch, once per folder
# anyone opened in Finder; there is nothing in them any checkout needs.
.DS_Store