yuzhenm/delm-720-trajectories

Paper | Project page | Codebase

DeLM trajectories on Terminal-Bench and DeepSWE

This dataset contains 720 agent trajectories from 240 DeLM runs on Terminal-Bench 4 and DeepSWE v1.1, using Astra xhigh and Opus 5.5 xhigh with n=2 and n=4 agents. Each of the eight settings includes ten tasks and three runs per task: 60 trajectories per n=2 setting and 120 per n=4 setting.

Trajectories include model interactions, tool calls and outputs, shared-context records, and available prompts and grading results. Failed and timed-out runs are included.

Dataset structure

README.md
archives/<setting>.zip    Trajectories grouped into eight settings

Settings are named <model>-<benchmark>-n<agents>: astra or opus55, tb (Terminal-Bench 4) or ds (DeepSWE v1.1), and n2 or n4.

Download and unzip a setting archive to access its trajectories. Inside each ZIP, agent files use this structure:

<setting>/<task>/R<run>/
├── run.json
└── agent-<number>/
    ├── native-trajectory.jsonl
    ├── trajectory.jsonl
    └── result.json

Native logs retain their Claude Code or Codex format; trajectory.jsonl contains the normalized event stream. Additional run and agent files contain available prompts, shared-context records, runtime information, and grading evidence.