| # VLMAuditor scripted trajectories, generator v2.1 (script3) | |
| Synthetic training trajectories for small-VLM auditors in WorldAuditBench's MCP agent setting, generated on our own procedurally built worlds (not benchmark scenes). | |
| - `script3_success_clean.jsonl`: 7,806 records (format `mcp-v1`: `tools` + `messages`), scripted explorer v2.1 (`MCPGEN_SCRIPT_V=2.1`), every record a clean success with its thoughts written; the 7 held-out evaluation worlds are excluded. | |
| - `images-script3-000..004.tar`: the 49,829 images these records need that are not in the earlier (private) snapshots of `kimperyang/vlmauditor-mcpgen-data`. Paths inside the tars are repo-relative (`out/mcpgen-20261023/...`): extract into the same IMAGE_ROOT as those snapshots. | |
| - `SHA256SUMS`: `sha256sum -c SHA256SUMS`. | |
| No benchmark tasks, rubrics or answers are included. | |