rlopd-data
收藏资源简介:
该数据集是`rlopd`项目的一部分,用于复现arXiv:2604.13016 §3.1中的受控对照实验,旨在比较on-policy distillation (OPD)与GRPO在相同步数下的表现。实验固定学生模型为Qwen3-1.7B-Base,教师模型有两种血统(Qwen3-4B-Base-GRPO和Qwen3-4B non-thinking),步数均为350。数据集包含三个实验组:`run_C1a`、`run_C1b`和`run_B2`,其中`run_B`因reward函数异常作废并已删除。数据内容包括每个实验组的模型检查点:merged HF格式(bf16权重+tokenizer,约3.3 GB)用于评测与发布,以及raw FSDP checkpoint(约22 GB)包含权重、优化器状态、RNG、调度器、dataloader位置,可用于续训。此外,还包含实验日志、配置以及reward判分器超时机制的迭代记录。该数据集适用于强化学习、知识蒸馏、模型训练与评估等相关研究。
This dataset is part of the `rlopd` project, used to reproduce the controlled comparison experiments in arXiv:2604.13016 §3.1, aiming to compare on-policy distillation (OPD) with GRPO under the same number of steps. The experiment fixes the student model as Qwen3-1.7B-Base, with two teacher model lineages (Qwen3-4B-Base-GRPO and Qwen3-4B non-thinking), both with 350 steps. The dataset includes three experimental groups: `run_C1a`, `run_C1b`, and `run_B2` (where `run_B` was discarded due to abnormal reward function and has been deleted). Each group contains model checkpoints: merged HF format (bf16 weights + tokenizer, ~3.3 GB) for evaluation and release, and raw FSDP checkpoints (~22 GB) including weights, optimizer states, RNG, scheduler, and dataloader positions, which can be used for continued training. Additionally, it includes experiment logs, configurations, and iterative records of the reward scorer timeout mechanism. This dataset is suitable for research on reinforcement learning, knowledge distillation, model training, and evaluation.





