---
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
pretty_name: Domyn Edge Teacher Code Mix
size_categories:
  - 1M<n<10M
tags:
  - domyn
  - domyn-edge
  - teacher
  - code
  - grpo
configs:
  - config_name: sft
    data_files:
      - split: train
        path: data/sft/train-*.parquet
      - split: test
        path: data/sft/test-*.parquet
  - config_name: rl
    data_files:
      - split: train
        path: data/rl/train-*.parquet
      - split: test
        path: data/rl/test-*.parquet
---

<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 Mix

> [!IMPORTANT]
> **Phased release.** Domyn Edge is a rolling release in three bundles. This dataset is part of the
> second bundle, together with the domain teachers and their post-training data. For an overview,
> see the [launch blog post](https://domyn.webflow.io/blog/domyn-edge-compact-model-open).

The training data of [Domyn Edge Teacher Code](https://huggingface.co/domyn/Domyn-Edge-Teacher-Code),
in two configs:

| Config | Stage | Train | Test |
|---|---|---:|---:|
| `sft` | Supervised fine-tuning, 64k context | 2,793,453 (38.48B tokens) | 147,130 (2.03B tokens) |
| `rl` | GRPO, unit-tested programming problems | 30,910 (19.90M tokens) | 1,069 (0.65M tokens) |

## `sft`: supervised fine-tuning mix

The code and software-engineering slices of the Domyn Edge SFT corpus. Tool-use data belongs to the
agentic teacher and math to the math teacher, so both have weight 0. Conversations longer than
65,536 tokens were dropped. 5% of the mix is held out as the test split. Conversations whose prompts overlap with the final evaluation benchmarks are excluded, to keep the mix usable for evaluation.

| Task | Weight | Sources |
|---|---:|---|
| Code | 1.0 | [Nemotron-SFT-Competitive-Programming-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Competitive-Programming-v2), [INTELLECT-3-SFT](https://huggingface.co/datasets/PrimeIntellect/INTELLECT-3-SFT) (`openreasoning_code`), [Dolci-Think-SFT-7B](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-7B), [Nemotron-SFT-CUDA-v1](https://huggingface.co/datasets/nvidia/Nemotron-SFT-CUDA-v1) |
| Software engineering | 1.0 | [Nemotron-SFT-SWE-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-SWE-v2), [Nemotron-SFT-SWE-v3](https://huggingface.co/datasets/nvidia/Nemotron-SFT-SWE-v3), [INTELLECT-3-SFT](https://huggingface.co/datasets/PrimeIntellect/INTELLECT-3-SFT) (`swe_swiss`) |

Released conversations and tokens per source, train and test splits together (licenses in short;
full terms in [ATTRIBUTION.md](ATTRIBUTION.md)):

<!-- table:teacher_code_sft -->
| Source | License | Conversations | Tokens |
|---|---|---:|---:|
| [INTELLECT-3-SFT](https://huggingface.co/datasets/PrimeIntellect/INTELLECT-3-SFT) | CC BY 4.0, MIT, Apache 2.0 (per subset) | 1,448,126 | 19.52B |
| [Nemotron-SFT-Competitive-Programming-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Competitive-Programming-v2) | CC BY 4.0, ODC-BY 1.0 | 692,261 | 14.01B |
| [Dolci-Think-SFT-7B](https://huggingface.co/datasets/allenai/Dolci-Think-SFT-7B) | ODC-BY 1.0 | 540,841 | 3.49B |
| [Nemotron-SFT-SWE-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-SWE-v2) | CC BY 4.0 | 251,494 | 3.33B |
| [Nemotron-SFT-SWE-v3](https://huggingface.co/datasets/nvidia/Nemotron-SFT-SWE-v3) | CC BY 4.0 | 5,643 | 142.0M |
| [Nemotron-SFT-CUDA-v1](https://huggingface.co/datasets/nvidia/Nemotron-SFT-CUDA-v1) | CC BY 4.0 | 2,218 | 21.4M |
| **Total** | | **2,940,583** | **40.51B** |
<!-- /table:teacher_code_sft -->

## `rl`: GRPO problems

Programming problems with unit tests. The reward is the fraction of tests the program in the
response passes, run in a sandbox. Each prompt carries a system message that asks for the final
solution in a single Python code block. Problems that overlap with the final evaluation benchmarks are excluded, to keep the config usable for evaluation.

| Source | License | Train rows | Train tokens | Test rows | Test tokens |
|---|---|---:|---:|---:|---:|
| [DeepCoder-Preview-Dataset](https://huggingface.co/datasets/agentica-org/DeepCoder-Preview-Dataset) | MIT | 20,352 | 11.84M | 939 | 0.55M |
| [Nemotron-3-Nano-RL-Training-Blend](https://huggingface.co/datasets/nvidia/Nemotron-3-Nano-RL-Training-Blend), code slice | CC BY-SA 4.0 | 10,558 | 8.06M | 130 | 0.10M |
| **Total** | | **30,910** | **19.90M** | **1,069** | **0.65M** |

Tokens are prompt tokens, as in the `token_length` column. For DeepCoder rows, `metadata.subset`
names the original collection (for example `codeforces`).

## How to use

```python
from datasets import load_dataset

sft = load_dataset("domyn/Domyn-Edge-Teacher-Code-Mix", "sft", split="train", streaming=True)
rl = load_dataset("domyn/Domyn-Edge-Teacher-Code-Mix", "rl", split="train")
```

## Columns

### `sft` columns

| Column | Type | Description |
|---|---|---|
| `messages` | list | The conversation. Each message has `role`, `content`, `reasoning_content` (assistant reasoning, if any), `tool_calls` (each with `id`, `type` and `function` {`name`, `arguments` as a JSON string}), and `name` and `tool_call_id` for tool results. |
| `tools` | string | JSON-encoded list of the tool schemas available in the conversation, or null. |
| `task` | string | Task family. |
| `source` | string | Upstream dataset. |
| `license` | string | License of the upstream record. |
| `metadata` | string | JSON-encoded upstream metadata (for example topic or generation model), or null. |
| `uuid` | string | Record identifier. |

### `rl` columns

The `rl` rows use the same canonical form as the `sft` rows, plus the reward data and the prompt length.

| Column | Type | Description |
|---|---|---|
| `messages` | list | The prompt: `role` and `content` of each message (the other message fields are null). |
| `tools` | string | Always null in this config. |
| `task` | string | Task family. |
| `source` | string | Upstream dataset. |
| `license` | string | License of the upstream record. |
| `metadata` | string | JSON-encoded upstream metadata. |
| `reward` | string | JSON-encoded reward data: `ground_truth` and the scoring `style`. Here `ground_truth.tests` is a JSON-encoded object with the test `inputs` and `outputs` and, for function-style problems, `fn_name`. |
| `token_length` | int | Prompt length in tokens. |
| `uuid` | string | Record identifier. |

To train with NeMo-Gym, as we did, map each row to the `code_gen_simple_agent` input:
`responses_create_params.input` = `messages`; `verifier_metadata.unit_tests` = the decoded
`reward.ground_truth.tests`; `agent_ref` = `{"type": "responses_api_agents", "name": "code_gen_simple_agent"}`.

## License

Our contributions to this dataset, including the curation and the mix composition, are released
under the Apache License 2.0. Upstream datasets keep their original license terms. Read
[ATTRIBUTION.md](ATTRIBUTION.md) before redistributing any part of this data.

## Citation

```bibtex
@misc{domyn_edge_teacher_code_mix_2026,
  title     = {Domyn Edge Teacher Code Mix},
  author    = {Domyn},
  year      = {2026},
  note      = {Technical report, forthcoming in November 2026.},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/domyn/Domyn-Edge-Teacher-Code-Mix}
}
```