---
license: cc-by-4.0
pretty_name: Blobfish Source Dataset Catalog
language:
  - en
tags:
  - dataset-discovery
  - mcp
  - tool-use
  - metadata
configs:
  - config_name: harbor
    data_files:
      - split: inventory
        path: harbor.jsonl.gz
  - config_name: huggingface
    data_files:
      - split: inventory
        path: huggingface.jsonl.gz
---

# Blobfish Source Dataset Catalog

**This is a metadata inventory, not an executable benchmark or a release of rebuilt worlds.**

Explore [the Blobfish dataset page](https://blobfish.ai/datasets) and
[existing executable benchmarks](https://blobfish.ai/benchmarks).

## Coverage

Exported 2026-09-04T21:50:06+00:00. Harbor: 295 inventory
records, including 294 listings in its
last complete refresh. Historical listings remain in the inventory.
Hugging Face: 1,032,990 unique inspected sources; 1,032,988
were observed through unfiltered creation-order pagination. Pass status:
**complete**, 1,033 pages. The
remaining rows came from a historical search cache.

The collector follows every public API continuation cursor. This is a rolling
traversal, not an atomic snapshot of the platform: sources can be created,
renamed, removed, or made private during collection. A complete pass means no
continuation cursor remained, not that every source was semantically audited.
`in_pagination_pass` distinguishes refreshed metadata from retained cache rows.
`manifest.json` contains file counts, SHA-256 digests, observation times, and
scan completion state. Unlike the website's bounded preview, these files contain
every inspected inventory row.

## Classification and reuse

- MCP/tool-use and professional-domain labels are heuristic metadata routing
  signals, not proof that a source exposes runnable MCP tools.
- `queued` requires public ungated access, a recognized open license, and MCP
  or domain relevance. It means candidate status, not a running build job.
- Unknown licenses and uncertain relevance remain in review. Restricted,
  gated, private, or disabled sources are not eligible for automatic reuse.
- Blobfish's own sources are marked `first-party` and excluded from duplicate
  adaptation. Ownership does not itself certify quality.
- `related_benchmark` is a domain reference, never evidence that this exact
  source has been rebuilt. Third-party adaptations remain `not-built`.

Only source identifiers, URLs, metadata facts, and Blobfish-authored annotations
are included. No upstream tasks, dataset contents, README prose, or assets are
redistributed. CC BY 4.0 applies to Blobfish's original annotations and catalog
presentation; it does not replace or grant rights under any upstream license.
Verify the pinned source and its terms before reuse.

## From a candidate to a benchmark

An adaptation requires independently authored tasks and assets, an isolated
stateful world with mocked CLI, web, API and MCP surfaces, passing oracle and
deterministic replays, rejected negative controls, cross-surface state parity,
and source-specific provenance. Publish a frozen website/Hugging Face/Harbor/
GitHub release, then measure models on that release before ranking them.

The existing [HubBench](https://blobfish.ai/benchmarks/hubbench) provides 104
tasks in 13 independently authored domain worlds. It is a separate executable
dataset, not a claim of per-row coverage for this inventory.

Collector and tests: [Blobfish dataset factory source](https://github.com/blobfishai/blobfishai/tree/main/benchmark/dataset_factory).