# Import evaluation results

Run this from the leaderboard repository after collecting the lm-eval JSON output:

```sh
node scripts/import-results.js --name "Model display name" --family "Model family" --license "Apache-2.0" path/to/results.json
```

The importer uses the model ID and total parameter count recorded by lm-eval. If those are absent or need correcting, pass `--model-id OWNER/MODEL` and `--params-m 135`. The display name and family default to the model ID's last component; the license defaults to `Not specified`.

Pass multiple JSON files after the options if a model's benchmarks were run separately. Every file must refer to the same model, and the combined files must include ARC-Easy, ARC-Challenge, GoldenSwag, and MMLU-redux `acc_norm`, and LAMBADA `bits_per_byte`. MMLU-redux continuation BPB is imported when available and is not used in the index. Limited or non-zero-shot runs, missing metrics, conflicting results, and existing model IDs are rejected. The importer records the input paths as the source, so keep those results files with the leaderboard.

Use `--data path/to/data.js` to write to a different copy of the leaderboard data. Run `node scripts/import-results.js --help` for the options.
