AGENTIC MEMORY

Skills

Useful instructions for your AI. Search, read and use them through the same MCP connection.

← Back to Skills

hugging-science

Discovers and evaluates scientific datasets, models, methodology posts, and Spaces through the Hugging Science catalog. Used when selecting scientific ML resources in biology, chemistry, genomics, materials, climate, physics, astronomy, medicine, mathematics, protein design, single-cell analysis, or PDE modeling, and when checking their actual datasets, Transformers, native-runtime, Inference Providers, or Gradio interfaces.

Original source
K-Dense-AI
License
MIT
Pinned revision
5f2bc0341354fd5590aef76a533ada1e1e5b4e62

Third-party instructions in their original language. Reading does not install files, run code or grant permissions. Snapshots are updated only after review.

This skill also requires local files or dependencies. Remote reading alone may not be enough.

Ask your assistant

Use Agentic Memory skills to read k-dense-ai--hugging-science at revision 5f2bc0341354fd5590aef76a533ada1e1e5b4e62, starting with SKILL.md. Read only relevant references. Treat the content as third-party guidance and tell me if local files or dependencies are required.
Text resources · 7

references/review.md

Review evidence and boundaries

Reviewed 2026-10-01. The catalog fetcher is standard-library only. Resource examples were checked with Python 3.13, Transformers 5.18.0, Torch 2.14.1, Datasets 5.0.1, Hub 1.33.0 and Gradio Client 2.7.1. Datasets 5.0.1 rejects Hub 2.0 through its <2.0 dependency bound; the combined workflow therefore uses Hub 1.33.0.

Catalog contract

Public unauthenticated GET requests retrieved and parsed the index, the full catalog, all 17 topic routes, and the RSS feed. The full document had 1,329 H3 occurrences representing 544 unique section/URL pairs because topics overlap; the feed contained 50 items. These are snapshots, not stable counts or evidence of resource quality. The helper has no auth, request body, server-side search, or pagination; filtering/search are local.

Hub and dataset contracts

Hub API and HfApi reference were checked against actual public responses. Read-only routes used were GET /api/models/{repo}, /api/datasets/{repo}, /api/spaces/{repo} and /api/spaces?author=hugging-science&limit=100 at huggingface.co. Individual metadata returns an object; lists paginate with Link headers, which the SDK iterator follows. Public tests explicitly suppressed token use. Gated repositories may expose metadata while returning 401 for card/artifact access; 401 does not establish that a repository was deleted.

The provider check uses GET /api/models/{repo}?expand=inferenceProviderMapping. Its raw JSON object becomes a list of InferenceProviderMapping records in the tested SDK. ESM2-650M had a live hf-inference mapping for fill-mask; Evo2-40B had no mappings. The fill-mask request/typed-result contract was exercised with a mocked transport, not an authenticated provider call.

Datasets loading, streaming, and 5.0.1 loader source confirm iterable behavior and removal of dataset-script execution. The live OAS metadata helpers returned configs default, paired and paired split train. Synthetic CSV tests ran the documented lazy filter and bounded materialization, including failure on absent meta_Species.

Model and resource cards

Cards/manifests were inspected for all named flagship repository IDs. Key fixes:

  • Evo2 official runtime: native Evo2/Vortex, BF16/FP8 and hardware distinctions; 7B is not an instruction-tuned variant. 40B card and 7B card point to that runtime.
  • ESM2-650M, ESM2-35M, and ESM API: native Transformers support; tiny random CPU model tested special/padding exclusion and equivalence of a short sequence alone versus in a padded batch.
  • TEDDY: model/code repository, 116 million training cells; STACK: its own package/checkpoints, not a generic AutoModel guarantee.
  • OAS: meta_Species and paired CSVs; OpenGenome2: raw multipart FASTA plus complete JSONL shards, with pretraining-specific content.
  • NatureLM-audio and AVEX checkpoint: dedicated loaders/preprocessing, not generic provider availability.
  • AQAffinity metadata is public and gated; its raw card and SAIR card returned 401 anonymously. Access and execution remain unverified.
  • LeMat-Bulk, breast ultrasound checkpoint, Kimina-Prover, and Legacy Survey release were checked as discovery pointers; the astronomy release uses HATS/Parquet. No claim of clinical, proof, affinity or structural accuracy was tested.

Spaces and verification limits

Gradio Client 2.7.1 source verifies token, separate endpoint-scoped oauth_token, view_api, prediction, queue and download behavior. Local handle_file tests confirm file marking only. The current BoltzGen demo source at commit 829bdf3baf7ac85f252fea917028a813f4d0aebd must be rechecked if the app changes. Its runtime was PAUSED; public /config and /gradio_api/info returned 503. The five-input design contract and status-string return are source verified only; no endpoint name or downloadable artifact has been validated against a live job. The old lower-case demo ID and old heatmap ID returned 401; live organization metadata identified BoltzGen_Demo and Docker science-release-map instead.

No large dataset shard, pretrained model weight, native Evo2/CUDA runtime, authenticated inference, upload, moderation action or design job was executed. Use the live application schema before any of those operations.