fuvty/mini-opd-deepmath

---
license: apache-2.0
pretty_name: Mini-OPD DeepMath
configs:
- config_name: opd
data_files:
- split: train
path: "opd/*.parquet"
- config_name: sft
data_files:
- split: train
path: "sft/*.parquet"
- config_name: sft_qwen3_4b
data_files:
- split: train
path: "sft_qwen3_4b/*.parquet"
- config_name: test
data_files:
- split: test
path: "test/*.parquet"
---

# Mini-OPD DeepMath

Private dataset package for Stage 0 / Stage 1 work in `mini-opd`.

Repo: `fuvty/mini-opd-deepmath`
Owner: `fuvty`

## Configs

### `opd`

Head of the DeepMath train split reserved for later on-policy distillation experiments.

- Split: `train`
- Rows: 53,000
- Columns: `question`, `ground_truth`, `source_index`
- Source indices: `0..52999`

### `sft`

Raw SFT candidate pool from the tail of the DeepMath train split before teacher filtering.

- Split: `train`
- Rows: 44,870
- Columns: `question`, `ground_truth`, `source_index`
- Source indices: `53000..97869`

### `sft_qwen3_4b`

Teacher-generated conversations kept from the Stage 0 rollout run with `Qwen/Qwen3-4B`.

- Split: `train`
- Rows: 35,256
- Columns: `question`, `ground_truth`, `source_index`, `messages`
- Source indices: subset of `53000..97869`

Important: this config contains only the retained correct conversations from the saved Stage 0 run. Incorrect teacher completions were not persisted in that run and therefore are not included in this release.

### `test`

Official DeepMath test split for validation/evaluation.

- Split: `test`
- Rows: 5,152
- Columns: `question`, `ground_truth`, `source_index`

## Source Dataset

- Source: `trl-lib/DeepMath-103K`
- Observed train rows in this environment: 97,870
- Observed test rows in this environment: 5,152

## Stage 0 Teacher Rollout Metadata

- Teacher model: `Qwen/Qwen3-4B`
- Sampling: `temperature=0.6`, `top_p=0.95`, `max_tokens=14336`
- Parallel rollout setting used in the run: `dp=8`
- Filter: keep only completions that pass `score_boxed_math_answer`
- Saved correct rate: 35,256 / 44,870 = 78.6%

## Notes

- `source_index` always refers to the index inside the original source split.
- `ground_truth` is the gold DeepMath answer, not the teacher completion.
- In `sft_qwen3_4b`, the full saved teacher conversation is stored in `messages` as a JSON string.