SKU AGT-4001 · Sold by External
AgentInstruct (AgentTuning)
Product specifications
| SKU | AGT-4001 |
|---|---|
| Data type | Agent traces |
| Volume | 1866 trajectories (across 6 task splits: os 195, db 538, alfworld 336, webshop 351, kg 324, mind2web 122) |
| Size on disk | ~8.04 MB uncompressed (dataset_size 8,042,511 bytes); ~1.26 MB Parquet download (download_size 1,255,385 bytes) |
| Format | Parquet (ShareGPT-style conversation records: conversations list + id) |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open dataset (HF) |
| Quality score | — |
| License | No license published by source |
AgentInstruct is the instruction-tuning dataset from the AgentTuning work by Tsinghua University (THUDM) and Zhipu AI. It contains 1,866 high-quality multi-turn interaction trajectories across six agent environments: Operating System (195), Database (538), ALFWorld household (336), WebShop (351), Knowledge Graph (324), and Mind2Web (122). Each trajectory is a ShareGPT-style conversation list where 'human' turns carry the environment/system prompt and observations and 'gpt' turns carry ReAct-style Thought/Action responses, with a per-turn boolean 'loss' flag marking which assistant turns are training targets. Trajectories were produced with GPT-4 (via Task Derivation and Self-Instruct) and retained only if they satisfied each task's strict reward/success criterion, so the released records are the filtered reference demonstrations rather than a runtime grader. The dataset is distributed as Parquet on the Hugging Face Hub (originally THUDM/AgentInstruct, now zai-org/AgentInstruct).