About

uncoolred bench is a UserBenchmark-style leaderboard for local LLM benchmarks at bench.uncool.red. Run benchmarks with benchmark_suite against vLLM, llama.cpp, TGI, or any OpenAI-compatible endpoint and submit to compare hardware, software stack, model, and recipe.

Components

• Astro static site — this site (leaderboard, runs, contributor profiles).
• Cloudflare Worker — POST /api/submissions validates the bundle (recipe, summary, metadata schema, handle, size, rate limit) and writes to D1 + R2.
• Cloudflare D1 + R2 — metadata in D1 SQLite; run artifacts (tarballs, llm-perf JSON, lm-eval results) in R2.

Submission flow

1. Run benchmarks locally with benchmark_suite.
2. bs submit <result_dir> --handle <your-handle> packages the result dir as a tarball and POSTs it.
3. Worker validates the bundle, persists metadata to D1, stores tarball in R2.
4. Site reads from the Worker’s API and renders pages.

Handle format

Handles match ^[a-z0-9][a-z0-9-]{1,38}$.

Privacy

v1 is read-only browsing + handle-attributed submissions. No GitHub OAuth yet (Phase B). Be honest about the numbers you publish.

Source on GitHub.