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
metadata
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.

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.

Sample usage

From the Github repository:

Download the prefix pool:

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:

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