---
pretty_name: DN-MOPD Training Prompts
license:
- mit
- cc-by-4.0
language:
- en
task_categories:
- text-generation
tags:
- on-policy-distillation
- multi-teacher-distillation
- reinforcement-learning
- math
- code
- instruction-following
size_categories:
- 10K<n<100K
configs:
- config_name: student
  data_files:
  - split: train
    path: student/train_3domain.jsonl
- config_name: teacher_math
  data_files:
  - split: train
    path: teacher/math_train.jsonl
- config_name: teacher_code
  data_files:
  - split: train
    path: teacher/code_train.jsonl
- config_name: teacher_if
  data_files:
  - split: train
    path: teacher/if_train.jsonl
---

# DN-MOPD Training Prompts

These are the exact prompt sets used in the Qwen3.5 experiments of
**Beyond Teacher Assignment: Domain-Normalized Multi-Teacher On-Policy Distillation** (DN-MOPD).

- **Project page:** https://lixin.ai/DN-MOPD
- **Code:** https://github.com/LiXin97/DN-MOPD
- **Paper:** [arXiv:2609.35347](https://arxiv.org/abs/2609.35347)
- **Models:** [DN-MOPD collection](https://huggingface.co/collections/XINLI1997/dn-mopd-6aba5f0df7bd8732d7205ed4): the GRPO teachers and the DN-MOPD and Label students trained on these prompts (21 checkpoints)

The code repository downloads these files into `data/` with the same paths.

## Files

| Config | File | Rows | Used by |
|---|---|---|---|
| `student` | `student/train_3domain.jsonl` | 2,700 (900 math, 900 code, 900 IF) | every student run: Label, DN-MOPD, single-teacher OPD, uniform pool, dynamic router, annealed injection, the controls, SeqKD-SFT generation |
| `teacher_math` | `teacher/math_train.jsonl` | 8,000 | GRPO math teachers |
| `teacher_code` | `teacher/code_train.jsonl` | 8,000 (917 MB, mostly test cases) | GRPO code teachers |
| `teacher_if` | `teacher/if_train.jsonl` | 8,000 | GRPO instruction-following (IF) teachers |

The teachers and students trained on these files are released as `XINLI1997/DN-MOPD-Qwen3.5-*` (for example
[DN-MOPD-Qwen3.5-9B](https://huggingface.co/XINLI1997/DN-MOPD-Qwen3.5-9B) and
[DN-MOPD-Qwen3.5-9B-teacher-math](https://huggingface.co/XINLI1997/DN-MOPD-Qwen3.5-9B-teacher-math)); the
[collection](https://huggingface.co/collections/XINLI1997/dn-mopd-6aba5f0df7bd8732d7205ed4) lists all 21.

`MANIFEST.json` lists the sha256, size, row count and source counts of every file.
`ids/student_math_code_ids.json` lists the 1,800 math and code prompts selected for the student set.

## Row schema

```json
{"prompt": [{"content": "...", "role": "user"}],
 "label": "...",
 "domain": "math",
 "teacher": "math",
 "metadata": {"domain": "math", "data_source": "DeepMath-103K", "src_index": 13934}}
```

| Field | Meaning |
|---|---|
| `prompt` | One user message. The chat template is applied with `enable_thinking=false`. In `teacher/if_train.jsonl` the prompt is a plain string, which the trainer wraps as one user message. |
| `label` | **math:** the reference answer, graded by a rule-based math verifier. **code:** a JSON string `{"inputs": [...], "outputs": [...]}` of stdin/stdout test cases. **IF:** a JSON list of verifiable constraints. |
| `domain` / `metadata.domain` | `math`, `code` or `ifeval`. It selects the verifier and the teacher, and it is the domain DN-MOPD normalizes over. |
| `teacher` | The teacher that scores the prompt under label routing; always equal to `domain`. |
| `metadata.data_source`, `metadata.src_index` | The source dataset, and the row's position in its source pool. |

## How the files were built

**Math pool** (57,046 prompts) comes from [DeepMath-103K](https://huggingface.co/datasets/zwhe99/DeepMath-103K) (rev. `5cf055d`).
- Questions are deduplicated by text in first-occurrence order.
- A question is kept if any of its occurrences has difficulty ≥ 6.
- Each prompt is followed by the line "Please reason step by step, and put your final answer within \boxed{}."

**Code pool** (25,276 prompts) is every code row of [Eurus-2-RL-Data](https://huggingface.co/datasets/PRIME-RL/Eurus-2-RL-Data) (rev. `9776b13`).
- The system message is dropped.
- Each user message is followed by the line "You need to think first then write the Python code."

**IF prompts** are 8,000 generated writing tasks.
- Each carries 1–3 verifiable constraints from a closed set of 19 types.
- Every constraint set is checked so that an empty answer fails it.
- No third-party text is used.

**Student set.**
- **Math and code:** 900 math and 900 code prompts, chosen before the Qwen3.5 experiments as prompts whose Qwen3-4B pass rate lay in [0.125, 0.875]. The pass rate was measured from 8 samples at temperature 1.0.
- **IF:** 900 of the IF prompts.
- The combined set was shuffled with seed 42.

**Teacher pools** start from the student rows of their domain and extend through the source pool to 8,000 rows, skipping duplicates. The IF teacher uses all 8,000 generated prompts.

`data/build/build_data.py` in the code repository rebuilds all four files byte for byte from the public sources. The one input it cannot recompute is the choice of the 1,800 student math/code prompts, which is published as ids.

## Licenses

| Rows | Source | License |
|---|---|---|
| math | [zwhe99/DeepMath-103K](https://huggingface.co/datasets/zwhe99/DeepMath-103K) | MIT |
| code | [PRIME-RL/Eurus-2-RL-Data](https://huggingface.co/datasets/PRIME-RL/Eurus-2-RL-Data) | MIT |
| IF (`ifeval`) | generated by the DN-MOPD authors | CC BY 4.0 |
| `MANIFEST.json`, `ids/` | DN-MOPD authors | CC BY 4.0 |

Eurus-2-RL-Data draws its code problems from four datasets:
- CodeContests (CC BY 4.0)
- TACO (Apache-2.0)
- APPS (MIT)
- Codeforces submissions (no license stated on its card)

The problem statements originate from competitive-programming sites, and their copyright stays with the original authors. See `DATA_LICENSES.md` for details.

Some code test cases contain strings that look like IP addresses or access keys. They come from programming problems about parsing such strings and are kept unchanged because the reward uses them.

## Citation

```bibtex
@article{li2026dnmopd,
  title   = {Beyond Teacher Assignment: Domain-Normalized Multi-Teacher On-Policy Distillation},
  author  = {Li, Xin and Jiang, Hao and Gao, Xin and Wang, Annan and Xie, Yuchen and Guo, Jinghao and Qu, Xingwei and Zhang, Yichi and Yuen, Chau},
  journal = {arXiv preprint arXiv:2609.35347},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.35347}
}
```

Please also cite the upstream datasets:
- DeepMath-103K (He et al., 2025)
- Eurus-2-RL-Data / PRIME (Cui et al., 2025)
- CodeContests (Li et al., 2022)
- TACO (Li et al., 2023)
- APPS (Hendrycks et al., 2021)