Apppublic

kz-transformers/kaz-llm-lb

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

12likes

App README

Kaz LLM Leaderboard

Benchmark results for LLMs on Kazakh multiple-choice tasks.

How it works

State lives in two private datasets, not in this repo. The data/ and m_data/ files committed here are only seeds; they are overwritten on boot.

Dataset	Mounted at	Holds
`kz-transformers/kaz-llm-lb-metainfo`	`data/`	`leaderboard.json`, the rendered board
`kz-transformers/s-openbench-eval`	`m_data/`	one JSON per evaluated model

On boot the app downloads both, aggregates every file under m_data/model_data/external/ into leaderboard.json, and renders it.

Submitting a results JSON uploads it to s-openbench-eval and flags the leaderboard as stale. A background job checks every 10 minutes and, only if a submission arrived, republishes leaderboard.json and restarts the Space so the new state is picked up at boot.

Layout

app.py                              UI, submission handling, scheduled refresh
src/envs.py                         repo ids, token, paths
src/leaderboard/build_leaderboard.py  dataset download, leaderboard DataFrame
src/display/utils.py                column definitions
src/display/about.py                intro and benchmark description text
src/display/css_html_js.py          custom CSS
src/radial/radial.py                per-model radar chart

Configuration

Requires one secret, set under Settings → Variables and secrets:

—H4_TOKEN — token with read/write access to the two datasets above.

Never log this value. Hardware should be cpu-upgrade.