Skip to content
 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Benchmark Radar

An evidence-first daily radar for newly released AI benchmarks, evaluation methods, datasets, leaderboards, and data-quality work.

Every day, GitHub Actions queries primary or structured sources, deduplicates records, classifies them with a transparent taxonomy, ranks them using explainable signals, and publishes a GitHub Issue. It is inspired by agents-radar, with sources and scoring redesigned for benchmark and AI-data research.

What it tracks

  • New AI/LLM benchmarks and challenge sets
  • Evaluation frameworks, judge models, safety/capability evals, and leaderboards
  • Public AI datasets, preference data, synthetic data, and data releases
  • Data contamination, leakage, provenance, deduplication, and annotation-quality work

Default sources:

Source Required secret Role
arXiv No Primary paper discovery
Hugging Face Hub No Dataset repository discovery
GitHub No in Actions Code and artifact discovery
OpenAlex OPENALEX_API_KEY Scholarly metadata enrichment
Brave Search BRAVE_API_KEY Web and lab-blog discovery

The report remains useful without optional secrets. Missing optional sources are shown as warnings in the source-health table instead of being silently ignored.

How ranking works

Each item receives four visible scores:

  • Relevance: matches against benchmark, evaluation, dataset, and data-quality taxonomy
  • Evidence: primary/structured source, authorship, and cross-source artifact evidence
  • Recency: time since publication or material update
  • Adoption: logarithmically scaled stars, downloads, likes, or citations

The default priority is:

0.40 relevance + 0.25 evidence + 0.20 recency + 0.15 adoption

This is triage, not scientific quality adjudication or endorsement.

Run locally

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[dev]'
benchmark-radar

Outputs:

  • out/report.md: the exact GitHub Issue body
  • out/items.json: machine-readable evidence and source-health snapshot

Run checks:

ruff check .
pytest -q

Configure

Edit config.yml to change the lookback, threshold, queries, taxonomy, and report size. Copy .env.example to .env only for local use; never commit credentials.

Optional repository secrets:

OPENALEX_API_KEY
BRAVE_API_KEY

The built-in GITHUB_TOKEN is used automatically in Actions.

Daily publishing

.github/workflows/daily-radar.yml runs at 12:15 UTC and can also be started manually. It:

  1. collects and renders with read-only repository permission;
  2. uploads the Markdown and JSON evidence for 30 days;
  3. creates or updates the issue for that UTC date;
  4. prevents duplicate daily issues by exact-title lookup.

The workflow needs repository Issues enabled. The labels daily-radar and automated must exist; they are created during initial repository setup.

Provenance and limitations

  • Every entry links to its discovered primary or structured record.
  • Optional-source failures are visible.
  • Reports can contain false positives; always inspect the source.
  • A repository update is not necessarily a new release.
  • Publication dates differ across preprints, code, datasets, and formal publications.
  • This system does not automatically create ANX-Bench events or research claims.

License

MIT

About

Daily evidence-first radar for AI benchmarks, evaluations, datasets, and data quality.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages