| license: mit | |
| language: | |
| - en | |
| tags: | |
| - reinforcement-learning | |
| - agents | |
| - embodied-ai | |
| - instruction-following | |
| - webshop | |
| - alfworld | |
| # SkillRL Reproduction Data | |
| This public dataset bundles the local data needed to reproduce the SkillRL | |
| training and environment setup. | |
| ## Contents | |
| - `SkillRL-SFT-Data/`: supervised fine-tuning instruction/output pairs for | |
| ALFWorld, WebShop, and Search. | |
| - `alfworld/`: extracted ALFWorld task definitions, logic files, and detector | |
| checkpoint. | |
| - `webshop/data/`: full and 1,000-product WebShop JSON datasets. | |
| - `webshop/search_engine/`: generated Pyserini/Lucene resources and indexes. | |
| The current WebShop search resources contain 100 documents in the `100` | |
| variant and 1,000 documents in the default, `1k`, and `100k` variants, matching | |
| the upstream setup script's default 1,000-product input. | |
| ## Source project | |
| See [aiming-lab/SkillRL](https://github.com/aiming-lab/SkillRL) for training and | |
| evaluation code and the original licenses and citations of the bundled | |
| benchmarks. | |