odysseus/services/hwfit/data

Runtime model catalogs

The shipped hf_models.json and mlx_community_models.json contain independently authored empty JSON lists ([]). They contain no copied upstream rows, descriptions, weights or metadata. A fresh offline installation has no catalog recommendations until user data has been populated.

Use Rescan in Cookbook while online (the models API accepts refresh_catalog=1). Existing discovery code fetches selected Hugging Face organization collections and MLX community collections into DATA_DIR/hwfit/: hf_collection_models.json and mlx_community_models.json. These runtime cache files retain source/fetch timestamps and model rows, with the existing 24-hour freshness policy. Forced refresh bypasses freshness, invalidates the merged in-memory catalog, and preserves the existing cache/network failure behavior. Previously populated caches can supply offline results; a failed cold refresh leaves an explicit empty-state message. Refresh does not write shipped lists.

Maintenance commands from the repository root:

python scripts/add_hwfit_models.py
python scripts/backfill_model_release_dates.py --dry-run
python scripts/import_from_vllm_recipes.py --dry-run

Their catalog path is DATA_DIR/hwfit/hf_models.json, merged before runtime collection caches. The add/import commands can initialize a missing runtime catalog; backfill requires one. These commands use Hub metadata and, for recipe import, vLLM recipe inputs. Runtime metadata is user data, not an approved redistributable snapshot. Publishing any populated catalog requires its own source/rights review; metadata, copied descriptions and model-weight licenses are distinct. Do not copy runtime results into these repository lists.

Tests use tests/fixtures/hwfit_publication_models.json through the scoped tests/hwfit_publication_fixtures.py fixture. The rows are independently authored synthetic test inputs. No production snapshot is required to test platform, quantization and GGUF behavior.