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/using-spaces.md

Using Hugging Science Spaces

A Space can be Gradio, Docker or static content. An organization listing does not mean every entry is an inference service or has a public API. Check sdk, runtime and the actual application before passing inputs. Public Spaces can sleep, fail, queue or require paid compute; they are not an availability or cost guarantee.

Inspect before calling

uv pip install 'gradio-client==2.7.1' 'huggingface-hub<2' python-dotenv

Metadata-only discovery (no inference or uploads):

from huggingface_hub import HfApi
api = HfApi(token=False)
for space in api.list_spaces(author="hugging-science", limit=100):
    print(space.id, space.sdk)
info = api.space_info("hugging-science/BoltzGen_Demo")
print(info.sdk, info.runtime.stage)

list_spaces is a paginated iterator; limit caps the number yielded, not a promise of a complete organization inventory. Names are case sensitive: the reviewed BoltzGen demo is hugging-science/BoltzGen_Demo, not boltzgen-demo.

When the selected Gradio runtime is available, inspect its schema:

import os
from dotenv import load_dotenv
load_dotenv()
from gradio_client import Client

# token=False suppresses implicit credentials for a public app.
# Use token=os.environ["HF_TOKEN"] only when authentication is needed and authorized.
client = Client("hugging-science/BoltzGen_Demo", token=False,
                analytics_enabled=False, httpx_kwargs={"timeout": 30})
schema = client.view_api(return_format="dict")
print(schema)

Gradio Client 2.7.1 uses token= (older examples may say hf_token=). token=None allows the locally saved/HF_TOKEN credential. Access authentication and oauth_token= are distinct: the latter lets an app act on your behalf and is only sent to endpoints declaring OAuth requirements. Do not provide an OAuth token for ordinary inference. Review the actual app/operator and data-transfer scope regardless of its organization. Follow existing user authorization; clarify only when the proposed upload or execution is outside that scope.

BoltzGen: source contract, runtime currently unverified

On 2026-10-01 the public Space repository resolved with runtime PAUSED, while /config and /gradio_api/info returned HTTP 503. Thus Client discovery and end-to-end inference were not validated. Do not substitute an invented /generate API.

The reviewed app source registers run_boltzgen with five inputs, in order:

  1. Design YAML file.
  2. Target PDB/CIF file referenced by that YAML.
  3. Protocol: protein-anything, peptide-anything or nanobody-anything.
  4. Number of candidate designs.
  5. Budget: final diversified design count, at most the number of candidates.

Budget is not an LLM token allowance. The run returns a status string; it does not directly return sequences or coordinates. Separate download actions exist. Their exact generated API names and return arity must be checked in view_api() when the runtime recovers. The reviewed source has shared /tmp/output paths and a download function returning two values for a one-output event; file isolation and download behavior need verification before relying on a public run.

After the live schema is confirmed, an illustrative call pattern is:

from gradio_client import handle_file

# Set this from the live schema; it is deliberately not a guessed endpoint.
api_name = confirmed_run_api_name
status = client.predict(
    handle_file("design.yaml"), handle_file("target.pdb"),
    "protein-anything", 2, 1, api_name=api_name,
)
print(status)

handle_file marks an input for upload. A local filename alone may be interpreted as text. It does not upload until the client call is submitted. Returned Gradio file components normally download to the client's local output directory; status text containing a server path is not itself a downloaded file. Use the actual output schema and verify artifacts belong to the current job.

Other reviewed organization entries

| Space | What the public metadata/source establishes | |---|---| | hugging-science/anatomy-of-boltzgen | Static educational site; no Gradio inference API | | hugging-science/dataset-quest | Gradio dataset discovery/submission app; submission is a write | | hugging-science/science-release-map | Docker visualization; old science-release-heatmap ID did not resolve publicly | | hugging-science/HuggingMod | Gradio moderation tooling, not scientific inference |

Do not retry a submitted design job blindly after a timeout: inspect its job/status first to avoid duplicate compute. Duplicating a Space provisions separate resources and may incur charges; it is a deployment action, not a read-only workaround. When no runtime/API is available, report that limitation and use the author's local workflow if it is feasible and within the task.

Sources: Gradio Client 2.7.1, file handling, HfApi Space methods.