simreadybench Claude Code and Codex runs (MuJoCo and Isaac)
Generation runs of the sim-ready benchmark with the official agent CLIs, Claude Code 2.1.322 (claude-opus-5-5) and Codex CLI 0.167.1 (gpt-6.1-sol): 1380 runs, each budget a fresh session, budgets 900/1800/2700 s, default reasoning effort (medium). The benchmark's items came in two splits; campaign in a run's run.json says which one it belongs to.
| split | campaign in run.json |
engine | items | runs |
|---|---|---|---|---|
| 1 | jiarui-mujoco |
MuJoCo | 51 | 2 x 153 = 306 |
| 1 | jiarui-isaac-a, jiarui-isaac-b |
Isaac Sim | 111 | 2 x 333 = 666 |
| 2 | jiarui-2 |
MuJoCo | 34 | 2 x 102 = 204 |
| 2 | jiarui-2 |
Isaac Sim | 34 | 2 x 102 = 204 |
The CLIs ran unmodified with their own tools (Claude Code without WebFetch; Codex with live web search). Model inference went through an OpenAI/Anthropic-compatible proxy instead of a login.
Layout: <track>/<engine>/<item>/<run_id>/ (track = scenes, objects or demos) with run.json, generation.jsonl (the CLI's JSON event log), submitted_prompt.txt, input/, output/. The model of a run is model_family in its run.json. Each split's campaign state (every job and the run it produced) is in campaigns/jiarui/state.json and campaigns/jiarui-2/state.json, merged from the hosts that ran it.
Not evaluated yet. Some host-specific strings and credentials are redacted. The Hub refuses paths containing /x00, so files named x00* are stored as _x00*.
The CodeBuddy runs (Kimi, GLM, DeepSeek, Hunyuan) of the same benchmark are in FlippyDora/simreadybench-codebuddy-runs.
- Downloads last month
- 184