---
license: cc-by-4.0
language:
- en
pretty_name: SENSUS Bench v1 (training data and evaluation kit)
task_categories:
- text-generation
tags:
- mathematics
- education
- tutoring
- synthetic
- supervised-fine-tuning
- evaluation
size_categories:
- 1K<n<10K
configs:
- config_name: training
  default: true
  data_files:
  - split: train
    path: training/train.jsonl
- config_name: evaluation
  data_files:
  - split: test
    path: evaluation/eval.jsonl
---

![SENSUS Bench v1.0: can a model read a learner's reasoning?](https://huggingface.co/datasets/Lorm-Labs/sensus-bench-v1/resolve/main/assets/banner.png)

# SENSUS Bench v1

Lorm Labs | 29 October 2026

SENSUS teaches and tests one tutoring skill: replying to a learner's written
reasoning about fractions and decimals (grades 4 and 5). A good reply credits
what the learner did correctly, names the likely error tentatively, explains
the idea briefly and ends with a question the learner can answer next.

| | |
|---|---|
| 🎮 **[Play Duel](https://huggingface.co/spaces/Lorm-Labs/sensus-bench-v1)** | Pick the better of two model replies and compare your choices with the graders |
| 📄 **[Paper](https://lorm.ai/paper.html)** | Teaching Models to Read and Respond to a Learner's Reasoning |
| 🌐 **[lorm.ai](https://lorm.ai)** | Results, the method and every evaluation case |

This repository holds two separate packages:

| Folder | What it is | Use it to |
|---|---|---|
| [`training/`](training/) | **SENSUS SFT v1.1**: 1,015 synthetic tutoring conversations, the recorded training recipe and step-by-step instructions | fine-tune your own model |
| [`evaluation/`](evaluation/) | **SENSUS Bench v1 Evaluation**: the 76 evaluation cases, the scripts that generate and grade replies, and every recorded reply and grade | score your model the same way the published results were scored |

No evaluation question appears in the training data. Keep the two apart: a
model trained on the evaluation cases would score higher without having learned
the skill.

All conversations and cases are synthetic. They contain no individual student
records.

## Train a model

```python
from datasets import load_dataset
training = load_dataset("Lorm-Labs/sensus-bench-v1", "training", split="train")
print(len(training))            # 1015
print(training[0]["messages"])  # system, user, assistant
```

Each row has one field, `messages`: the tutor instruction, the question with
what the learner did and said, and the target tutor reply. The full recipe
(LoRA rank 32, 8 epochs, learning rate 0.0003) and how to reproduce the
recorded runs are in [`training/TRAINING.md`](training/TRAINING.md) and
[`training/recipe.json`](training/recipe.json). The dataset card is
[`training/README.md`](training/README.md).

## Evaluate a model

```python
from datasets import load_dataset
cases = load_dataset("Lorm-Labs/sensus-bench-v1", "evaluation", split="test")
print(len(cases))               # 76
```

To score a model, download the `evaluation/` folder and run:

```
python3 sample.py --model <your-model> --out replies.jsonl
python3 grade.py --replies replies.jsonl --out grades.jsonl
```

`grade.py` uses two model graders and needs your own API keys. The published
models' replies and both graders' marks are in `evaluation/recorded/`. Details
are in [`evaluation/README.md`](evaluation/README.md).

## Check the files

Each folder has a `verify.py` that checks every file against its published
fingerprints, offline:

```
cd training && python3 verify.py
cd evaluation && python3 verify.py
```

## Results and paper

On these 76 cases, Qwen3.8-27B fine-tuned on SENSUS SFT scores 0.825 on average
over two training runs, against 0.290 for the same model before training and
0.525 for the highest-scoring general-purpose model tested. The method, results
and limits are in the technical report, *Teaching Models to Read and Respond to
a Learner's Reasoning*: https://lorm.ai/paper.html

## License

The data in both folders is released under the Creative Commons Attribution 4.0
International License (CC BY 4.0); see `training/LICENSE` and
`evaluation/LICENSE`. The scripts in both folders are released under the MIT License;
see `training/LICENSE-CODE` and `evaluation/LICENSE-CODE`.

## Citation

If you use the benchmark or the results, please cite the paper:

```bibtex
@techreport{wang2026sensus,
  title       = {Teaching Models to Read and Respond to a Learner's Reasoning: SENSUS Bench and an Open SFT Dataset for Grade 4 and 5 Fraction and Decimal Errors},
  author      = {Wang, Karen},
  institution = {Georgetown University},
  year        = {2026},
  month       = oct,
  url         = {https://lorm.ai/paper.html}
}
```

If you use the training data, please also cite the dataset:

```bibtex
@misc{lorm2026sensussft,
  title        = {{SENSUS SFT}: Synthetic Tutoring Conversations for Supervised Fine-Tuning},
  author       = {{Lorm Labs}},
  year         = {2026},
  version      = {1.1},
  howpublished = {\url{https://lorm.ai/datasets.html}},
  note         = {Dataset released under CC BY 4.0}
}
```

The evaluation kit's citation is in `evaluation/README.md`.

## Contact

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