Karl28/RSIPO-SFT-Data

RSIPO Stage-1 Skill SFT Data

The exact Stage-1 training and validation sets used to initialize the six RSIPO runs reported in the paper, together with the underlying policy trajectories and GLM-5.2 trajectory-to-summary/skill annotations.

• Paper: RSIPO: Recursive Self-Improving Policy Optimization for Agentic Reinforcement Learning via Skill-Policy Co-Evolution
• Code: https://github.com/zfkarl/RSIPO
• SFT initialization checkpoints: linked below
• Final 150-step RL checkpoints: the checkpoint table in the code repository

Configurations

Config	Base model	Benchmark	Train	Validation	Valid annotations	SFT checkpoint
qwen25-3b-alfworld	Qwen2.5-3B-Instruct	ALFWorld	3,595	399	3,994	Karl28/RSIPO-SFT-Qwen2.5-3B-Instruct-ALFWorld
qwen25-7b-alfworld	Qwen2.5-7B-Instruct	ALFWorld	3,595	400	3,995	Karl28/RSIPO-SFT-Qwen2.5-7B-Instruct-ALFWorld
qwen3-1.7b-alfworld	Qwen3-1.7B	ALFWorld	3,590	399	3,989	Karl28/RSIPO-SFT-Qwen3-1.7B-ALFWorld
qwen25-3b-search	Qwen2.5-3B-Instruct	Search-based QA	3,598	400	3,998	Karl28/RSIPO-SFT-Qwen2.5-3B-Instruct-Search
qwen25-7b-search	Qwen2.5-7B-Instruct	Search-based QA	3,587	399	3,986	Karl28/RSIPO-SFT-Qwen2.5-7B-Instruct-Search
qwen3-1.7b-search	Qwen3-1.7B	Search-based QA	3,589	399	3,988	Karl28/RSIPO-SFT-Qwen3-1.7B-Search

The train and validation Parquets are byte-for-byte copies of the files consumed by Stage-1 SFT. Their hashes are in SOURCE_SHA256.json. Each model collected 8 policy rollouts on each of 500 training tasks; GLM-5.2 then wrote an episode summary and reusable episode skill for every trajectory that parsed successfully. The released split is the exact approximately 90/10 split used by training.

Load the SFT splits

from datasets import load_dataset

data = load_dataset("Karl28/RSIPO-SFT-Data", "qwen25-7b-alfworld")
print(data["train"][0]["prompt"])
print(data["train"][0]["response"])

The Parquet columns are:

• all configurations: prompt, response, skill_id, task_id, task_type, source_success, source_num_steps, parse_ok;
• search configurations additionally include row_index, data_source, source_final_task_score and analysis_prompt_version.

Trajectories and annotations

Each configuration directory also contains:

File	Contents
sampled_tasks.jsonl.gz	The 500 tasks sampled for rollout collection
trajectories.jsonl.gz	8 base-policy trajectories per task, including actions/observations and outcome
annotations.jsonl.gz	GLM-5.2 analysis prompt, raw output, parsed episode_summary and episode_skill
sft_records.jsonl.gz	All valid prompt/response SFT records before the train/validation split
metrics.json	Counts and aggregate statistics from data preparation
metadata.json	Configuration-level generation and SFT metadata

For example:

import gzip
import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "Karl28/RSIPO-SFT-Data",
    "qwen25-7b-alfworld/annotations.jsonl.gz",
    repo_type="dataset",
)
with gzip.open(path, "rt") as f:
    first_annotation = json.loads(next(f))
print(first_annotation["episode_summary"])
print(first_annotation["episode_skill"])

Data generation and SFT

The pipeline follows SEED and is reproduced in scripts/sft/ of the code repository:

1. sample 500 training tasks;
2. collect 8 trajectories per task from the corresponding base model (Qwen3 runs in non-thinking mode);
3. use GLM-5.2 to write a trajectory summary and reusable skill;
4. discard outputs that do not parse into the required JSON structure;
5. fine-tune for 3 epochs with learning rate 5e-6 and global batch size 8.

The six task selections are recorded in each configuration's sampled_tasks.jsonl.gz. In brief:

• qwen25-3b-alfworld: 84/83 examples from each of the six ALFWorld task types.
• qwen25-7b-alfworld: 84/83 examples from each of the six ALFWorld task types.
• qwen3-1.7b-alfworld: 84/83 examples from each of the six ALFWorld task types.
• qwen25-3b-search: 500 questions from NQ, TriviaQA, PopQA-like, HotpotQA, 2WikiMultiHopQA and MuSiQue.
• qwen25-7b-search: 100 questions each from NQ, TriviaQA, HotpotQA, 2WikiMultiHopQA and MuSiQue.
• qwen3-1.7b-search: 500 questions stratified across NQ, TriviaQA, PopQA-like, HotpotQA, 2WikiMultiHopQA and MuSiQue.

Licensing

This release contains derived examples from ALFWorld and public QA datasets distributed through Search-R1 and FlashRAG. It is marked license: other because the upstream datasets use different licenses. Users must comply with the licenses and terms of ALFWorld, NQ, TriviaQA, PopQA, HotpotQA, 2WikiMultiHopQA, MuSiQue, Search-R1 and FlashRAG as applicable. The GLM-5.2 annotations are provided for research and reproducibility.

Citation

@misc{zhang2026rsipo,
      title={RSIPO: Recursive Self-Improving Policy Optimization for Agentic Reinforcement Learning via Skill-Policy Co-Evolution},
      author={Fan Zhang and Hao Wu and Jinyang Wu and Hongyu Qu and Zheng Lian and Xiao Luo and Yefeng Zheng and Bo Zhang and Haoxuan Li and Pheng-Ann Heng},
      year={2026},
      eprint={2610.15772},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.15772},
}