---
license: cc-by-4.0
language:
- en
pretty_name: SENSUS Bench v1 Evaluation
task_categories:
- text-generation
tags:
- mathematics
- education
- tutoring
- synthetic
- evaluation
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: eval.jsonl
---

# SENSUS Bench v1 Evaluation

Lorm Labs | Evaluation kit v1.0 | 26 October 2026

## Summary

The 76 evaluation cases behind the SENSUS Bench v1.0 results, with the
scripts to generate and grade replies and the recorded replies and grades of
the published models. Use it with SENSUS SFT v1.1 to check whether a model you
fine-tune reproduces the recorded scores. None of these questions appear in
the training file.

## Files

- `eval.jsonl`: the 76 evaluation cases.
- `sample.py`: generates one reply per case from an OpenAI-compatible endpoint.
- `grade.py`: grades replies with the two-grader checklist method.
- `recorded/models.json`: the published models and their recorded scores.
- `recorded/replies.jsonl`: every recorded reply with both graders' marks.
- `LICENSE`: CC BY 4.0 for the data. `LICENSE-CODE`: MIT for the scripts.
- `verify.py`, `manifest.json`, `SHA256SUMS`: release fingerprints and an offline check.

## Data fields (`eval.jsonl`)

| Field | Content |
|---|---|
| `id` | `e001` to `e076`. |
| `standard` | Common Core State Standards for Mathematics code. |
| `messages` | The `system` and `user` messages given to the tutor model, in the training format. |
| `criteria` | Checklist lines a good reply meets; the first line says the reply states nothing false. |
| `critical` | Serious errors; marking any one scores the reply 0. |
| `reference` | A reply that meets the checklist, for reference only. |

## How to evaluate a model

```
python3 verify.py
python3 grade.py --check-recorded          # offline: reproduces the table below
python3 sample.py --model <your-model> --out replies.jsonl
python3 grade.py --replies replies.jsonl --out grades.jsonl
```

`sample.py` uses temperature 0, at most 900 output tokens and thinking off, the
recorded settings. `grade.py` calls two graders (`claude-opus-5-5` and
`gpt-6.1-sol` at medium reasoning effort) and needs `ANTHROPIC_API_KEY` and
`OPENAI_API_KEY`. A checklist line counts only if both graders mark it met; a
serious error marked by either grader scores the reply 0. Grading 76 replies
makes 152 grader calls. If the recorded grader versions are unavailable,
report the graders you used: scores from different graders are not directly
comparable.

## Recorded scores

| Model | Mean score | Passes (all lines met) |
|---|---|---|
| Qwen3.8-27B (Lorm) | 0.838 | 24 |
| Qwen3-14B (Lorm) | 0.720 | 14 |
| Qwen3-8B (Lorm) | 0.626 | 7 |
| GPT-6 Sol | 0.525 | 0 |
| Opus 5.5 | 0.461 | 0 |
| Nemotron 3 Ultra (NVIDIA) | 0.356 | 0 |
| Kimi K3 | 0.331 | 0 |
| Qwen3.8-27B (untuned) | 0.290 | 0 |

Two training runs of each fine-tuning recipe differed by up to 0.03, so a
reproduction should land near, not exactly on, these values.

## Limitations

- These cases are public. A model trained on them would score higher
  without being better, so they are not a hidden test.
- Scores are model-graded checklist measurements of reply quality on 76 cases.
  They are not measures of student learning.
- Coverage matches the training data: grade 4 fractions and grade 5 place
  value and decimals, in English.

## License

Data (`eval.jsonl`, `recorded/`): Creative Commons Attribution 4.0
International (CC BY 4.0). Scripts (`sample.py`, `grade.py`, `verify.py`): MIT.
Recorded replies of third-party models are included as evaluation outputs; their
use may also be subject to the terms of those models' providers.

## Citation

```bibtex
@misc{lorm2026sensusbench,
  title        = {{SENSUS Bench} v1 Evaluation},
  author       = {{Lorm Labs}},
  year         = {2026},
  version      = {1.0},
  howpublished = {\url{https://lorm.ai/datasets.html}},
  note         = {Data under CC BY 4.0; code under MIT}
}
```

## Contact

Use the contact form at https://lorm.ai/contact.html