diversemate/mdl-catalog

mdl catalog

catalog.sqlite is the snapshot mdl find reads: the Hugging Face model tree, as one SQLite file, rebuilt nightly.

mdl catalog pull          # fetch this file (ETag, so a no-op when unchanged)
mdl catalog stats
mdl catalog tree Qwen/Qwen3-8B
mdl find                  # the best model your machine can run

It is not a dataset to train on. It is an index, so a local tool can answer "what can this machine run, and how well" without crawling the hub itself.

What is in it

table	one row per	notable columns
nodes	model	id, official, parent, relation, arch, params, ctx, license, downloads, likes, tags, datasets
ggufs	GGUF file	repo, node, quant, file, size, shards, downloads
evals	reported score	node, benchmark, task, value, verified, source
meta	-	built_at, schema_version, counts

nodes is a forest: a base model, the official post-trains under it, and the fine-tunes and merges under those, with parent/relation edges taken from the hub's own model tree. ggufs links each quantized repo back to the model it quantizes. evals holds what model cards and model-index blocks report, which is a claim, not a measurement.

How it is built

A crawler walks the tracked families and follows base_model:finetune|merge and base_model:quantized edges, keeping models that clear a noise gate (500 downloads in 30 days, or 20 likes), and reads each GGUF repo's file list for sizes and quant names. No weights are downloaded. Source: mdl_fit/catalog.py, built by .github/workflows/catalog.yml.

Reading it without mdl

import sqlite3
db = sqlite3.connect("catalog.sqlite")
db.row_factory = sqlite3.Row
for r in db.execute("""SELECT n.id, g.quant, g.size FROM ggufs g
                       JOIN nodes n ON n.id = g.node
                       WHERE n.id = ? ORDER BY g.size""",
                    ("Qwen/Qwen3-8B",)):
    print(r["id"], r["quant"], r["size"])

Metadata is mirrored from the hub and carries whatever licenses and errors the source repos carry; the MIT license above covers the index itself.