| --- |
| pretty_name: MODELOME |
| tags: |
| - machine-learning |
| - model-catalog |
| - model-metadata |
| - provenance |
| - research |
| --- |
| |
| # MODELOME v0.1.0 |
|
|
| MODELOME is a provenance-preserving catalog of model, method, and technique entries with source-attributed identifiers and resource references. It is a metadata dataset; it contains no model weights or media. |
|
|
| ## Snapshot |
|
|
| This bounded snapshot has 44 shards and 26,082 entry rows, comprising 25,816 distinct exact entry IDs and 266 repeated rows across shards. Shard counts are summed without global deduplication; repeated IDs remain present in their source shards. This is partial coverage, not an exhaustive inventory. |
|
|
| The `release-index.json` records shard prefixes, counts, source observations, and integrity digests. The core shard replaces the v10 core with a newer Hugging Face continuation: 101 pages and 10,100 records seen, with upstream total unavailable and a resumable cursor. Source checkpoints describe the observed coverage. In particular, TensorFlow Hub reached cursor 475 of 2,371 eligible paths and remains partial, with 1,896 paths unvisited. The `overlap-report.json` records integrity and cross-shard identity checks: there are no exact-ID name conflicts, no normalized-name conflicts for repeated IDs, and no shared `(member.source, member.source_record_id)` pairs across shards. |
|
|
| Forty exact IDs repeat between the architecture-audio and Torchaudio shards. These are source-corroborating rows with consistent names; they are retained in both shards. ESPnet's `validated_by_espnet` status is preserved: 160 rows are `true` and two are `false`. The false rows remain historical source-listed identities with declared references and do not assert current validation or operational weights. Twelve source-declared `same_artifact` links remain in the language shard; no global merge is performed. |
|
|
| ## Files and layout |
|
|
| Each shard has the same bundle layout beneath its prefix. The root shard uses these paths directly; additional shards are under their indexed `sources/<prefix>/` paths. |
|
|
| - `entries/entries.jsonl` contains one JSON object per entry, including its exact ID, canonical name, aliases, identifiers, tags, source members, resources, releases, citations, and declared model relations. |
| - `metadata/*.parquet` contains normalized relational tables for models, artifacts, identifiers, aliases, releases, and source links. |
| - `manifest.json` and the entry manifest record counts and SHA-256 checksums for each shard. |
| - `seeds.jsonl` preserves the reusable entry seeds used to form the catalog. |
|
|
| Source declarations and references remain attached to their provenance. Source-level terms and licenses continue to apply; users should follow the cited sources for their materials. |
|
|
