Blame for README.md

---
library_name: transformers
license: apache-2.0
pipeline_tag: text-generation
base_model: domyn/Domyn-Edge-SFT
datasets:
  - domyn/Domyn-Edge-Teacher-Code-Mix
language:
  - en
tags:
  - domyn
  - domyn-edge
  - teacher
  - code
  - grpo
  - fully-open
---

<p align="center">
  <picture>
    <source media="(prefers-color-scheme: dark)" srcset="https://cdn.prod.website-files.com/682dcb35e7fd7a313cae803a/6835cb0fb6bd22a7cba7d645_domyn-logo-payoff-primary-white.svg">
    <img alt="Domyn" src="https://cdn.prod.website-files.com/682dcb35e7fd7a313cae803a/6835ca86622d317630fa3861_domyn-logo-primary.svg" width="400">
  </picture>
</p>

# Domyn Edge Teacher Code

> [!IMPORTANT]
> **Phased release.** Domyn Edge is a rolling release in three bundles. This teacher is part of the
> second bundle, together with the other domain teachers and their post-training data. See
> [Open resources](#open-resources) for everything released so far.

> [!Note]
> This repository holds the **code teacher** of Domyn Edge: Domyn Edge SFT after a code and
> software-engineering supervised pass and GRPO on unit-tested programming problems. It is one of
> the domain teachers that score the student's rollouts token by token in the multi-teacher
> on-policy distillation (MOPD) stage of our post-training protocol. It is released for research and reproducibility; it is not a
> general-purpose assistant.

## Table of contents

1. [Model summary](#model-summary)
2. [How to use](#how-to-use)
3. [Evaluation](#evaluation)
4. [Limitations](#limitations)
5. [Open resources](#open-resources)
6. [License](#license)
7. [Citation](#citation)

## Model summary

Domyn Edge Teacher Code is a 4B-parameter dense transformer with the same architecture and tokenizer
as [Domyn Edge SFT](https://huggingface.co/domyn/Domyn-Edge-SFT). It was trained in two stages:

1. **Code SFT**: one epoch at 64k context on competitive programming, general code, CUDA and
   repository-level software-engineering data.
2. **GRPO**: asynchronous group relative policy optimization on programming problems, rewarded by
   the fraction of unit tests the generated program passes.

The weights are stored in bfloat16.

## How to use

The model loads and serves exactly like Domyn Edge SFT; see the
[Domyn Edge SFT card](https://huggingface.co/domyn/Domyn-Edge-SFT#how-to-use) for the recommended
environment, Transformers and vLLM instructions.

**Chat template.** The repository ships the Domyn Edge SFT chat template unchanged; it is the
template used during GRPO training and in the evaluation below. It opens the system turn with a
fixed identity line ("You are Domyn Edge, a model trained by Domyn.").

**Output format.** The model reasons first and writes its final solution in a single fenced Python
code block at the end of the response. Responses to hard problems often run past 30k tokens, so
leave room in `max_new_tokens`.

## Evaluation

### Teacher vs student on the general evaluation suite

Scores of this teacher next to those of the student, [Domyn Edge
SFT](https://huggingface.co/domyn/Domyn-Edge-SFT), with one benchmark per area, to check what the
teacher keeps outside its training domain. All scores are percentages.

| Group | Benchmark | Metric | Domyn Edge SFT | Teacher Code | Δ |
|---|---|---|---:|---:|---:|
| Math | AIME 2026 | avg@10 | 61.3 | **66.3** | +5.0 |
| Multilingual math | MGSM (de, en, es, fr, ru) | accuracy | 85.8 | **87.5** | +1.7 |
| Knowledge and reasoning | GPQA Diamond | pass@1 | 56.3 | **65.4** | +9.1 |
| Instruction following | IFBench | prompt strict pass@1 | 24.1 | **26.0** | +1.9 |
| Function calling | BFCL v4, without web search (overall) | pass@1 | 36.5 | **37.7** | +1.2 |
|  | BFCL v4 (non-live) | pass@1 | 51.9 | **55.8** | +3.9 |
|  | BFCL v4 (live) | pass@1 | **75.9** | 75.4 | -0.5 |
|  | BFCL v4 (multi-turn) | pass@1 | 35.9 | **36.0** | +0.1 |
|  | BFCL v4 (memory) | pass@1 | 27.3 | **31.4** | +4.1 |
| Agentic | SWE-bench Verified | pass@1 | **33.6** | 25.0 | -8.6 |
| Code | LiveCodeBench (full set)\* | pass@1 | 73.0 | **74.0** | +1.0 |

_\*LiveCodeBench (full set) is all 1,055 problems of LiveCodeBench v6, released between May 2023 and
April 2025._

<details>
<summary>Evaluation settings</summary>

- Served with vLLM in bfloat16, with the chat template shipped in this repository.
- Thinking on. Sampling: temperature 1.0, top-p 0.95, top-k 20, min-p 0, no presence or
  repetition penalty.
</details>

### Code evaluation suite

| Benchmark | Metric | Teacher Code | Qwen3.5 4B | Δ |
|---|---|---:|---:|---:|
| HumanEval | pass@1 | **92.1** | 89.6 | +2.5 |
| MBPP (sanitized) | pass@1 | **82.9** | 72.0 | +10.9 |
| CRUXEval-O | pass@1 | 72.1 | **84.1** | -12.0 |
| CRUXEval-I | pass@1 | 81.4 | **89.1** | -7.7 |
| LiveBench coding | pass@1 | 64.2 | **72.1** | -7.9 |
| LiveCodeBench (Qwen set)\* | pass@1 | **53.1** | 47.3 | +5.8 |
| BigCodeBench | pass@1 | 33.0 | **46.9** | -13.9 |

_\*LiveCodeBench (Qwen set) is the 131 problems of LiveCodeBench v6 released from February to April 2025,
the subset selected by the 2025-02 to 2025-05 window that Qwen uses to report LiveCodeBench. These
recent problems are harder than the full set of 1,055 problems used in the teacher vs student table,
so the scores are lower._

<details>
<summary>Evaluation settings</summary>

- Served with vLLM in bfloat16 at 64k context, thinking on.
- HumanEval, MBPP, CRUXEval and LiveBench coding: temperature 1.0, top-p 0.95, top-k 20, min-p 0, no
  presence or repetition penalty.
- LiveCodeBench and BigCodeBench: temperature 0.6, top-p 0.95, top-k 20, min-p 0, no presence or repetition
  penalty.
- [Qwen3.5 4B](https://huggingface.co/Qwen/Qwen3.5-4B) is evaluated in the same settings, using its
  own chat template.

</details>

## Limitations

**A teacher, not an assistant.** This model is one of the domain teachers of multi-teacher on-policy
distillation (MOPD), the post-training stage in which the Domyn Edge student generates its own
responses and each teacher scores them token by token in its own domain.

**Long generations.** On hard problems the model can reason for tens of thousands of tokens, and a
small share of responses degenerate when they run far past 64k tokens. Cap the generation length.

**Safety work is incomplete.** No dedicated safety tuning or red teaming was done on this model.
Run generated code in a sandbox.

## Open resources

- **This teacher**: [`domyn/Domyn-Edge-Teacher-Code`](https://huggingface.co/domyn/Domyn-Edge-Teacher-Code)
- **Its data**: [`domyn/Domyn-Edge-Teacher-Code-Mix`](https://huggingface.co/datasets/domyn/Domyn-Edge-Teacher-Code-Mix) (SFT and RL splits)
- **Student SFT model**: [`domyn/Domyn-Edge-SFT`](https://huggingface.co/domyn/Domyn-Edge-SFT), and its mix [`domyn/Domyn-Edge-SFT-Mix`](https://huggingface.co/datasets/domyn/Domyn-Edge-SFT-Mix)
- **Collection**: [Domyn Edge](https://huggingface.co/collections/domyn/domyn-edge) on Hugging Face

## License

This model is released under [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0).

## Citation

```bibtex
@misc{domynedge2026,
  title  = {Domyn Edge},
  author = {Domyn},
  year   = {2026},
  note   = {Technical report, forthcoming in November 2026.},
}
```