math-cold-start / README.md
nielsr's picture
nielsr HF Staff
Add dataset card with paper link, task category and sample usage
6e05549 verified
|
Raw
History Blame
1.5 kB
---
task_categories:
- text-generation
language:
- en
---
# ReOPD Prefix Pool
This repository contains a prefix pool dataset from the paper [Multi-Turn On-Policy Distillation with Prefix Replay](https://huggingface.co/papers/2607.04763).
ReOPD (Replayed-Prefix On-Policy Distillation) distills a teacher agent into a student **without environment interaction during student training**: the student acts on prefixes replayed from pre-collected teacher trajectories, while the teacher supplies dense per-step supervision. The prefix pool is built from the teacher's RL rollouts with a step-decaying sampling schedule (`p_t ∝ κ^t`, κ = 0.6), favoring early, low-shift prefixes. It matches or improves online OPD accuracy with **zero tool calls** and **4–9× faster rollouts**.
- Project page: https://baohaoliao.github.io/ReOPD/
- Code: https://github.com/BaohaoLiao/ReOPD
- Hugging Face collection: https://huggingface.co/collections/baohao/reopd
## Sample usage
From the [Github repository](https://github.com/BaohaoLiao/ReOPD):
Download the prefix pool:
```bash
hf download baohao/Math_Qwen3-8B_SFT-RL_Prefix --repo-type dataset --local-dir /data/pool
```
Then use it as `PROMPT_DATA` in the ReOPD distillation script:
```bash
TEACHER=qwen3-8b STUDENT=qwen3-4b \
TEACHER_MODEL_PATH=/ckpt/teacher_hf \
HF_CHECKPOINT=/ckpt/student_hf \
REF_LOAD=/ckpt/student_torch_dist \
PROMPT_DATA=/data/pool/train.jsonl \
SAVE_DIR=/runs/reopd_8b_to_4b \
bash scripts/train/reopd_math.sh
```