---
title: README
emoji: 🧮
colorFrom: blue
colorTo: gray
sdk: static
pinned: false
---

![Lorm Labs: at the frontier of AI, for human understanding](https://huggingface.co/spaces/Lorm-Labs/README/resolve/main/banner.png)

**Lorm Labs** explores how advances in intelligence can expand what people understand, discover and solve. We publish open benchmarks and training data, and every release ships with its data, its evaluation and every recorded reply and grade.

### Latest release · SENSUS Bench v1.0

When a learner makes a mistake with fractions or decimals, the reasoning they write down usually shows why. SENSUS tests whether a model can use that evidence: name the likely error, say it as likely and not certain, credit what is correct, explain the idea and end with a next step.

| Model | SENSUS Bench score |
|---|---|
| **Qwen3.8-27B + SENSUS SFT** (mean of two runs) | **0.825** |
| GPT-6 Sol (highest-scoring general-purpose model) | 0.525 |
| Qwen3.8-27B, untuned | 0.290 |

| | |
|---|---|
| 🎮 **[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 |
| 📦 **[Dataset](https://huggingface.co/datasets/Lorm-Labs/sensus-bench-v1)** | 1,015 synthetic training conversations and the 76-case evaluation kit, CC BY 4.0 |
| 📄 **[Paper](https://lorm.ai/paper.html)** | Teaching Models to Read and Respond to a Learner's Reasoning |

[lorm.ai](https://lorm.ai) · [Contact](https://lorm.ai/contact.html)