parallell commited on
Commit
7948366
·
verified ·
1 Parent(s): 0902b96

Upload final MOPD-Com labeled training data and dataset card

Browse files
Files changed (2) hide show
  1. README.md +102 -0
  2. data/train_prompt2048.parquet +3 -0
README.md ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: MOPD-Com Training Data
3
+ size_categories:
4
+ - 10K<n<100K
5
+ task_categories:
6
+ - text-generation
7
+ tags:
8
+ - mopd
9
+ - knowledge-distillation
10
+ - math
11
+ - code
12
+ - instruction-following
13
+ configs:
14
+ - config_name: default
15
+ data_files:
16
+ - split: train
17
+ path: data/train_prompt2048.parquet
18
+ ---
19
+
20
+ # MOPD-Com Training Data
21
+
22
+ 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 It is a source-labeled, preprocessed dataset, not a claim of an exact reproduction of an unpublished paper preprocessing pipeline.
23
+
24
+ ## Domain composition
25
+
26
+ | Routing label (`extra_info.opd_teacher`) | Examples |
27
+ | --- | ---: |
28
+ | `math` | 25,000 |
29
+ | `code` | 25,000 |
30
+ | `instruction_following` | 15,994 |
31
+ | Total | 65,994 |
32
+
33
+ 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.
34
+
35
+ ## Sources and sampling
36
+
37
+ - Math: [icemoon28/MOPD-Training-Data](https://huggingface.co/datasets/icemoon28/MOPD-Training-Data), `DeepMath-103K/train_filtered_level6.parquet` (57,046 source rows).
38
+ - Code: the same dataset, `Eurus/code_train.parquet` (25,276 source rows).
39
+ - 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.
40
+
41
+ Pinned icemoon28 revision: `96e88d579cb1ee68c9183c52e5d133f58cf370bf`.
42
+ Pinned HIR revision: `2a95f69eb56cc47edc16a45f939cde479673a4cb`.
43
+ Per-domain shuffle seeds: math 42, code 43, instruction following 44; final mixture shuffle seed 42. No deduplication was applied.
44
+
45
+ ## Prompt construction
46
+
47
+ Math and code message contents are preserved. HIR prompts include the original question followed by all provided criteria, using this format:
48
+
49
+ ```text
50
+ {original question}
51
+
52
+ The output must follow the following rules:
53
+ 1. {first criterion}
54
+ 2. {second criterion}
55
+ ...
56
+ ```
57
+
58
+ 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.
59
+
60
+ 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.
61
+
62
+ ## Schema
63
+
64
+ - `data_source`: source identifier.
65
+ - `prompt`: list of messages with `role` and `content`, suitable for the project's chat template interface.
66
+ - `ability`: ability/domain metadata.
67
+ - `reward_model`: struct containing `style` and string `ground_truth`. Original reference/verifier content is preserved; JSON content is serialized as a string where needed.
68
+ - `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.
69
+
70
+ 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.
71
+
72
+ ## Usage
73
+
74
+ ```python
75
+ from datasets import load_dataset
76
+ train = load_dataset('parallell/MOPD-Com-Training-Data', split='train')
77
+ ```
78
+
79
+ To download the original prepared Parquet file:
80
+
81
+ ```python
82
+ from huggingface_hub import hf_hub_download
83
+ path = hf_hub_download(
84
+ repo_id='parallell/MOPD-Com-Training-Data',
85
+ repo_type='dataset',
86
+ filename='data/train_prompt2048.parquet',
87
+ local_dir='mopd_training_data',
88
+ )
89
+ ```
90
+
91
+ 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.
92
+
93
+ ## Integrity
94
+
95
+ - File: `data/train_prompt2048.parquet`
96
+ - Bytes: 785742653
97
+ - Rows: 65994
98
+ - SHA256: `96d7fef2f0b4a72e7282b76d7a40d55413e7b4d84387d1c417e1ef08e612a7ec`
99
+
100
+ ## Upstream terms and attribution
101
+
102
+ 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.
data/train_prompt2048.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:96d7fef2f0b4a72e7282b76d7a40d55413e7b4d84387d1c417e1ef08e612a7ec
3
+ size 785742653