File size: 4,607 Bytes
7948366
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
---
pretty_name: MOPD-Com Training Data
size_categories:
- 10K<n<100K
task_categories:
- text-generation
tags:
- mopd
- knowledge-distillation
- math
- code
- instruction-following
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train_prompt2048.parquet
---

# MOPD-Com Training Data

This repository contains the final labeled training mixture prepared for [MOPD-Com](https://github.com/Lorsense/MOPD-Com). It contains **65,994 training examples** in one Parquet file. It is a source-labeled, preprocessed dataset, not a claim of an exact reproduction of an unpublished paper preprocessing pipeline.

## Domain composition

| Routing label (`extra_info.opd_teacher`) | Examples |
| --- | ---: |
| `math` | 25,000 |
| `code` | 25,000 |
| `instruction_following` | 15,994 |
| Total | 65,994 |

The original sampled mixture had 25,000 math, 25,000 code, and 16,000 instruction-following examples. Six sampled HIR examples exceeded the 2,048-token prompt limit and were removed. The remaining mixture order is preserved. Domain labels were assigned from source provenance; no model was called for labeling.

## Sources and sampling

- Math: [icemoon28/MOPD-Training-Data](https://huggingface.co/datasets/icemoon28/MOPD-Training-Data), `DeepMath-103K/train_filtered_level6.parquet` (57,046 source rows).
- Code: the same dataset, `Eurus/code_train.parquet` (25,276 source rows).
- Instruction following: [sastpg/HIR-16K](https://huggingface.co/datasets/sastpg/HIR-16K), `HIR_trainv1.jsonl` (16,968 source rows). The instruction-following subset of icemoon28/MOPD-Training-Data was not used.

Pinned icemoon28 revision: `96e88d579cb1ee68c9183c52e5d133f58cf370bf`.
Pinned HIR revision: `2a95f69eb56cc47edc16a45f939cde479673a4cb`.
Per-domain shuffle seeds: math 42, code 43, instruction following 44; final mixture shuffle seed 42. No deduplication was applied.

## Prompt construction

Math and code message contents are preserved. HIR prompts include the original question followed by all provided criteria, using this format:

```text
{original question}

The output must follow the following rules:
1. {first criterion}
2. {second criterion}
...
```

This corresponds to a layout supported by the [official HIR helper](https://github.com/sastpg/HIR/blob/ffd6919ca69a9dafa0c9977be745d209856c8a39/HIR/verl/trainer/ppo/ray_trainer.py). No system prompt is included in any example. No target assistant answer is appended to the prompt.

Prompt lengths were measured using the Qwen/Qwen3-1.7B tokenizer with the chat template, `add_generation_prompt=True`, and `enable_thinking=False`. Every retained example has a prompt length of at most 2,048 tokens. Other tokenizers or chat template settings may produce different lengths.

## Schema

- `data_source`: source identifier.
- `prompt`: list of messages with `role` and `content`, suitable for the project's chat template interface.
- `ability`: ability/domain metadata.
- `reward_model`: struct containing `style` and string `ground_truth`. Original reference/verifier content is preserved; JSON content is serialized as a string where needed.
- `extra_info`: routing label `opd_teacher`, unique index, source dataset/revision/file/row and original metadata, original HIR prompt and criteria, prompt construction metadata, prompt SHA256, prompt token length, and length/system-prompt flags.

The pure on-policy distillation training configuration uses generated rollouts and teacher/student probabilities. The preserved ground truth is not an assistant completion inserted into the input prompt.

## Usage

```python
from datasets import load_dataset
train = load_dataset('parallell/MOPD-Com-Training-Data', split='train')
```

To download the original prepared Parquet file:

```python
from huggingface_hub import hf_hub_download
path = hf_hub_download(
    repo_id='parallell/MOPD-Com-Training-Data',
    repo_type='dataset',
    filename='data/train_prompt2048.parquet',
    local_dir='mopd_training_data',
)
```

For local MOPD-Com training, point the training data path to the downloaded Parquet file. This repository contains only the train split; no held-out validation split is provided.

## Integrity

- File: `data/train_prompt2048.parquet`
- Bytes: 785742653
- Rows: 65994
- SHA256: `96d7fef2f0b4a72e7282b76d7a40d55413e7b4d84387d1c417e1ef08e612a7ec`

## Upstream terms and attribution

This is a processed mixture of the linked upstream datasets. Consult the upstream dataset cards and original sources for applicable licenses and use conditions. This card does not assign a new blanket license or replace upstream terms.