README.md

metadata

license: apache-2.0
task_categories:
  - question-answering
  - text-generation
language:
  - en
pretty_name: APEX Agents evaluation trajectories

APEX Agents Evaluation Trajectories

This dataset contains the lightweight trajectory artifacts for one full Archipelago / APEX Agents evaluation run.

Run

• Run id: 20260717_174205_250s_c29
• Benchmark: Mercor Archipelago / APEX Agents
• Agent config: react_toolbelt_agent
• Max steps: 250
• Concurrency: 29
• Model: local Qwen3.6-27B-SFT-v10-tau-style-v4-ckpt340 served by SGLang
• Tasks: 480

Score Summary

• Strict Pass@1: 65 / 480 = 0.1354
• Partial average: 0.2822
• No-grade counted as zero: 25
• Infra failures: 0

Domain Summary

Domain	Tasks	Strict Pass	Strict Pass@1	Partial Avg
Investment Banking	160	18	0.1125	0.1646
Law	160	20	0.1250	0.3582
Management Consulting	160	27	0.1688	0.3237

Files

• data/trajectories.jsonl: one JSON object per task, including task metadata, trajectory, grades, verifiers, initial messages, and run score fields.
• raw_tasks/<task_id>/: per-task JSON artifacts copied from the evaluator output.
• logs/: run manifest, task ids, aggregate results JSONL, and runner log.
• summary.json: machine-readable summary.

Large environment snapshots and filesystem archives were intentionally excluded from this upload because the original evaluator output directory is hundreds of GB. This dataset is meant for trajectory inspection and score analysis, not complete environment replay.

Common API key/token patterns were redacted during staging.