This is the Supervised Fine-Tuning (SFT) dataset used in the paper SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning.
SkillRL-SFT-Data provides instruction-output pairs for training base agent policies on three interactive decision-making environments: ALFWorld, WebShop, and Search. Each example contains a structured instruction with retrieved skill context from the hierarchical SkillBank and the corresponding expert action output.
| Config | Environment | Examples | Description |
|---|---|---|---|
alfworld | ALFWorld | 7,486 | Embodied household tasks (pick & place, clean, heat, cool, examine, etc.) |
webshop | WebShop | 2,553 | Web-based shopping navigation tasks |
search | Search | 1,214 | Multi-step web search QA tasks (NQ, TriviaQA, PopQA, HotpotQA, 2Wiki, MuSiQue, Bamboogle) |
Total: 11,253 instruction-output pairs.
Each example contains two fields:
instruction: A detailed prompt including the task goal, retrieved relevant experience from the hierarchical SkillBank (general principles, task-specific skills, and mistakes to avoid), current environment observation, and admissible actions.output: The expected agent response, consisting of a step-by-step reasoning process wrapped in <think> tags followed by the selected action in <action> tags.from datasets import load_dataset
# Load a specific config
alfworld_data = load_dataset("Jianwen/SkillRL-SFT-Data", "alfworld")
webshop_data = load_dataset("Jianwen/SkillRL-SFT-Data", "webshop")
search_data = load_dataset("Jianwen/SkillRL-SFT-Data", "search")
| Environment | SFT Checkpoint | RL Checkpoint |
|---|---|---|
| ALFWorld | Alfworld-7B-SFT | Alfworld-7B-RL |
| WebShop | Webshop-7B-SFT | Webshop-7B-RL |
| Search | Search-7B-SFT | Search-7B-RL |
All models are fine-tuned from Qwen2.5-7B-Instruct. SFT checkpoints are trained on this dataset; RL checkpoints are further optimized via recursive skill-augmented reinforcement learning.
@article{xia2026skillrl,
title={SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning},
author={Xia, Peng and Chen, Jianwen and Wang, Hanyang and Liu, Jiaqi and Zeng, Kaide and Wang, Yu and Han, Siwei and Zhou, Yiyang and Zhao, Xujiang and Chen, Haifeng and others},
journal={arXiv preprint arXiv:2602.08234},
year={2026}
}
This is the Supervised Fine-Tuning (SFT) dataset used in the paper SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning.
SkillRL-SFT-Data provides instruction-output pairs for training base agent policies on three interactive decision-making environments: ALFWorld, WebShop, and Search. Each example contains a structured instruction with retrieved skill context from the hierarchical SkillBank and the corresponding expert action output.
| Config | Environment | Examples | Description |
|---|---|---|---|
alfworld | ALFWorld | 7,486 | Embodied household tasks (pick & place, clean, heat, cool, examine, etc.) |
webshop | WebShop | 2,553 | Web-based shopping navigation tasks |
search | Search | 1,214 | Multi-step web search QA tasks (NQ, TriviaQA, PopQA, HotpotQA, 2Wiki, MuSiQue, Bamboogle) |
Total: 11,253 instruction-output pairs.
Each example contains two fields:
instruction: A detailed prompt including the task goal, retrieved relevant experience from the hierarchical SkillBank (general principles, task-specific skills, and mistakes to avoid), current environment observation, and admissible actions.output: The expected agent response, consisting of a step-by-step reasoning process wrapped in <think> tags followed by the selected action in <action> tags.from datasets import load_dataset
# Load a specific config
alfworld_data = load_dataset("Jianwen/SkillRL-SFT-Data", "alfworld")
webshop_data = load_dataset("Jianwen/SkillRL-SFT-Data", "webshop")
search_data = load_dataset("Jianwen/SkillRL-SFT-Data", "search")
| Environment | SFT Checkpoint | RL Checkpoint |
|---|---|---|
| ALFWorld | Alfworld-7B-SFT | Alfworld-7B-RL |
| WebShop | Webshop-7B-SFT | Webshop-7B-RL |
| Search | Search-7B-SFT | Search-7B-RL |
All models are fine-tuned from Qwen2.5-7B-Instruct. SFT checkpoints are trained on this dataset; RL checkpoints are further optimized via recursive skill-augmented reinforcement learning.
@article{xia2026skillrl,
title={SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning},
author={Xia, Peng and Chen, Jianwen and Wang, Hanyang and Liu, Jiaqi and Zeng, Kaide and Wang, Yu and Han, Siwei and Zhou, Yiyang and Zhao, Xujiang and Chen, Haifeng and others},
journal={arXiv preprint arXiv:2602.08234},
year={2026}
}