early-experience
收藏资源简介:
本数据集是论文《Agent Learning via Early Experience》的复现数据,旨在支持智能体通过早期经验进行学习的研究。它覆盖了八个智能体交互环境:ALFWorld、WebShop、BFCL、TravelPlanner、TextCraft、AppWorld、ScienceWorld和WebArena。针对每个环境,数据提供了三种训练范式所需的数据拆分:模仿学习(IL,对应expert拆分)、自我反思(SR,需要expert和reflection拆分)以及隐式世界建模(IWM,对应iwm拆分)。所有数据样本均采用ShareGPT对话格式,包含system、user和assistant消息字段。数据规模因环境而异,例如ALFWorld的expert、iwm和reflection拆分分别包含21,335、192,008和21,331条样本;WebShop则分别为15,678、92,662和6,216条。值得注意的是,WebArena环境仅提供iwm和reflection拆分,缺少expert拆分。部分环境(如AppWorld、WebShop、WebArena)中的数据包含由基准测试生成的合成虚构个人信息(如假邮箱、地址、URL),并非真实数据。该数据集适用于强化学习、模仿学习、世界模型构建等智能体学习任务的研究与模型微调。
This dataset is a reproduction of data from the paper Agent Learning via Early Experience, aimed at supporting research on agent learning through early experiences. It covers eight different agent interaction environments: ALFWorld, WebShop, BFCL, TravelPlanner, TextCraft, AppWorld, ScienceWorld, and WebArena. For each environment, the data provides splits required for three training paradigms: imitation learning (IL, corresponding to the expert split), self-reflection (SR, requiring both expert and reflection splits), and implicit world modeling (IWM, corresponding to the iwm split). All data samples are in ShareGPT dialogue format, including system, user, and assistant message fields. The data scale varies by environment; for example, ALFWorlds expert, iwm, and reflection splits contain 21,335, 192,008, and 21,331 samples respectively, while WebShop has 15,678, 92,662, and 6,216 samples. Notably, the WebArena environment only provides iwm and reflection splits, lacking the expert split. Data in some environments (e.g., AppWorld, WebShop, WebArena) contains synthetic fictional personal information (such as fake emails, addresses, URLs) generated by benchmark tests and is not real data. This dataset is suitable for research and model fine-tuning in agent learning tasks such as reinforcement learning, imitation learning, and world model construction.
数据集概览
数据集名称: Early Experience (Reproduction Data)
许可证: MIT
语言: 英语
任务类别: 强化学习 (Reinforcement Learning)
标签: agent, early-experience, imitation-learning
描述
本数据集是用于复现论文《Agent Learning via Early Experience》的监督微调数据,覆盖8个智能体环境。每个环境提供三种训练范式的数据:
- IL (Imitation Learning): 使用
expert数据 - SR (Self-Reflection): 使用
expert+reflection数据 - IWM (Implicit World Modeling): 使用
iwm(世界模型) +expert数据
数据集结构
数据集按环境(Config)和数据类型(split)组织,所有记录采用 ShareGPT 格式(包含 system / user / assistant 消息)。
环境与数据切分
| 环境 | expert | iwm | reflection |
|---|---|---|---|
| alfworld | 21,335 | 192,008 | 21,331 |
| webshop | 15,678 | 92,662 | 6,216 |
| bfcl | 1,224 | 8,677 | 752 |
| travelplanner | 1,370 | 52,754 | 1,226 |
| textcraft | 374 | 16,200 | 2,975 |
| appworld | 931 | 10,067 | 931 |
| scienceworld | 2,038 | 158,795 | 35,065 |
| webarena | — | 144,894 | 33,727 |
注意: WebArena 仅提供 iwm 和 reflection 数据,无 expert 切分。
使用方式
python from datasets import load_dataset
加载指定环境与切分的数据
ds = load_dataset("osunlp/early-experience", "alfworld", split="reflection")
复现结果
以下结果展示了早期体验方法(IWM、SR)相比模仿学习(IL)的性能提升(成功率或各基准的原始指标)。除 AppWorld 使用 Qwen2.5-14B-Instruct 外,其余环境均使用 Qwen2.5-7B-Instruct。
| 环境 | 模型 | IL | IWM | Δ vs IL | SR | Δ vs IL |
|---|---|---|---|---|---|---|
| ALFWorld | Qwen2.5-7B-Instruct | 58.6% | 68.8% | +10.2% | 74.2% | +15.6% |
| WebShop | Qwen2.5-7B-Instruct | 39.1% | 48.4% | +9.3% | 55.5% | +16.4% |
| BFCL | Qwen2.5-7B-Instruct | 44.9% | 51.1% | +6.2% | 49.6% | +4.7% |
| TravelPlanner | Qwen2.5-7B-Instruct | 25.0% | 29.4% | +4.4% | 30.0% | +5.0% |
| TextCraft | Qwen2.5-7B-Instruct | 73.0% | 76.8% | +3.9% | 74.5% | +1.5% |
| ScienceWorld | Qwen2.5-7B-Instruct | 65.4% | 68.6% | +3.2% | 66.0% | +0.6% |
| AppWorld | Qwen2.5-14B-Instruct | 43.7% | 59.7% | +16.0% | 51.0% | +7.3% |
注意: WebArena 仅作为数据发布,未在上述评估中体现。
其他说明
- AppWorld 使用 Qwen2.5-14B-Instruct,其余环境使用 Qwen2.5-7B-Instruct。
- WebArena 未发布
expert切分,其观察内容为可访问性树,可能包含基准测试自托管环境的 URL。 - AppWorld、WebShop、WebArena 包含由基准测试生成的虚构个人数据(如虚假邮箱、密码、地址、URL),并非真实数据。




