Real numbers from real codebases. Every entry was produced by running neuralmind benchmark . --json on the submitter's own repository and PR'd into this dataset. CI validates every submission against a schema. NeuralMind never uploads anything automatically — these are deliberate, auditable contributions.
Each dot is one submission. X-axis is the graph node count (proxy for codebase size); Y-axis is the average reduction ratio vs loading the whole repo. The pattern to watch for: does the ratio hold up as repos get bigger?
Hover any dot for project name, language, and exact numbers.
Reduction ratios per language. Bars show the average across submissions; the count overlay is how many entries back each number — so you can tell apart "this is solid" from "n=1, take with salt".
Sortable. The verification_command column is the exact CLI invocation the submitter ran — copy it, run it on your own repo, and the numbers should reproduce. That's what makes this table auditable rather than promotional.
| Project | Language | Nodes | Reduction | Wakeup tok | Query tok | Model | Submitted by | Date | Notes |
|---|---|---|---|---|---|---|---|---|---|
| Loading submissions… | |||||||||
The dataset gets stronger the more repos it covers — especially across different languages and sizes. To contribute:
cd /path/to/your-repo
neuralmind build .
neuralmind benchmark . --json > my-bench.jsonThen open a PR adding one entry to docs/community-benchmarks.json (schema enforced; see community-benchmarks.schema.json for required fields). CI will validate the entry on submission; this dashboard updates automatically once merged.