Datasets:

---

FrontierCS

/

Frontier-CS

like 6

Follow

FrontierCS 7

We maintain inspect_evals, an open-source collection of evaluations for the Inspect AI framework, and we use this dataset in our FrontierCS eval. Our CI validates dataset paths via the HF dataset-viewer endpoint:

https://datasets-server.huggingface.co/is-valid?dataset=FrontierCS/Frontier-CS

It currently returns all-false ({"preview": false, "viewer": false, "search": false, "filter": false, "statistics": false}), and the dataset viewer page shows:

Cannot get the split names for the config 'default' of the dataset.
Exception:    HfHubHTTPError
Message:      504 Server Error: Gateway Time-out for url:
              https://huggingface.co/api/datasets/FrontierCS/Frontier-CS/paths-info/...

The dataset itself loads fine; load_dataset("FrontierCS/Frontier-CS", split="test") works and returns records with problem_id, category, statement, config. The traceback shows the viewer is timing out while globbing the whole repo to infer data-file patterns (get_data_patterns → resolve_pattern → fs.glob → get_paths_info). That's likely because the repo contains ~3.9k files (test data, solutions, scripts), so listing all paths exceeds the gateway timeout, even though the actual tabular data is a single data/test-00000-of-00001.json.

Adding an explicit configs: block to the dataset card's YAML front matter should fix this — it points the resolver straight at the data file instead of globbing the entire tree:

configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*.json

This should let the viewer resolve splits without the expensive full-repo path listing. Happy to help if useful, thanks!