jm-vis/modelroom-catalog

Open Model Catalog

A small table of the models their publishers call current: text, vision, embedding, speech recognition, image and video generation. Not every build on the Hub. One row per model, with the publisher page that says it is current and the date that was checked.

252 models · 57 publisher accounts watched · one JSON file (260 KB) · schema 2, and schemas 3 and 4 with findings · CC BY 4.0

Three files, three schemas

The dataset holds the same table three times, one file per schema, so that each release of the reader finds the file it knows:

File	Schema	Read by
catalog-v2.json	2	release 1.0.0; unchanged in form, it carries no findings
catalog-v3.json	3	release 1.1.0; the rows of schema 2, plus gguf_checks and setup_note (a note without vllm_args)
catalog-v4.json	4	releases that read the findings and the switches below; schema 3, plus vllm_args in a setup_note

The files are made from one working table and published together; the published files are compared byte for byte with the repository every day. A row of schema 3 or 4 without findings reads as "not checked".

What a row carries

Field	Meaning
hf_repo, publisher, family, kind	the repository, its account, the model family, and one of eight kinds: text, vision, embedding, speech-recognition, speech-synthesis, image-generation, video-generation, audio
parameters_b, license	parameter count in billions, the license the card states
latest, latest_source, latest_checked	whether the publisher calls it current, the page that says so, and the date that was checked
successor, successor_source, successor_checked	the model that replaced it, when the publisher named one
ollama_base, ollama_tag	the Ollama library entry, where one exists
gguf_repos	GGUF builds of five packagers: Unsloth, bartowski, mradermacher, LM Studio, ggml
hub_pipeline_tag, kind_source	the task the Hub states, and where the kind came from (the tag, the base model, the config, or a decision)
source, added, note	the page the row was read from, the day it was added, a short note where a decision needed one

Schema 3 (catalog-v3.json) adds two findings to a row, and schema 4 (catalog-v4.json) one field to the second. Each says what was compared and when, and nothing else.

Field	Meaning
gguf_checks	one entry per GGUF package, bound to its repository, revision and weight file (or first shard): the architecture the file states, compared with the LLM_ARCH_NAMES block of the src/llama-arch.cpp file of one llama.cpp commit. present or absent, the commit, the page, the day, and where the architecture was read (the Hub's field for the one file it describes, else the first bytes of the file)
setup_note	a pointer to the model card of the publisher's repository at one revision: whether the card holds a vllm serve start block for the model, and the name of a package the card states as required for inference, where its wording is explicit
setup_note.vllm_args	schema 4 only: the switches that the start block names for this model, of a closed list (--reasoning-parser, --tool-call-parser, --enable-auto-tool-choice, --kv-cache-dtype), each with a value of letters, digits, ., _ and -; nothing else of a start block (no path, no --trust-remote-code, no memory or GPU switch) is kept

What these findings do not say: present is no promise that a runtime loads the file, and absent is a statement about one package against one commit of one source, never about a runtime version. A package or a card without an entry was not checked or could not be read; a failed run adds no absent and removes no entry. A setup_note does not say the card lists everything a model needs, and none_recognized says only that the rule of the scan found nothing. The scan reads the cards and file headers; it runs nothing from them.

The file also lists the publishers (account, display name, region such as EU, CN, US, KR, JP) and the packagers whose GGUF builds are recorded.

How it stays current

Every morning a scan reads the 57 publisher accounts and the trending page of each task and writes a proposal; a part of the catalog rows per day also gets its GGUF packages and model card checked (the findings above). Nothing goes live on its own: the maintainer approves every row by number, rejected accounts stay rejected, and the published file is compared byte for byte with the repository every day.

The limit is the maintainer. The scan finds what the Hub shows; whether a row is right is one person's judgement, with the page that supports it in the row.

Use it

from datasets import load_dataset

catalog = load_dataset("jm-vis/modelroom-catalog")

Or read the file directly:

import json, urllib.request

url = "https://huggingface.co/datasets/jm-vis/modelroom-catalog/resolve/main/catalog-v2.json"
catalog = json.load(urllib.request.urlopen(url))
current = [m for m in catalog["models"] if m["latest"]]

The files with the findings are catalog-v3.json and catalog-v4.json, at the same address.

The first consumer is ModelRoom, which reads the catalog at the start of each run to answer "which model fits this machine".

Something missing?

Open an issue in the ModelRoom repository naming the account and the page that names its current model. A wrong row: the page that says otherwise.

License and attribution

The data is published under CC BY 4.0. Use it, copy it, change it, build on it, commercially too; name the source:

modelroom catalog, jm-vis, CC BY 4.0, https://huggingface.co/datasets/jm-vis/modelroom-catalog

The models listed keep their own licenses; the license field states what each model card says.