| --- |
| license: apache-2.0 |
| tags: [terminal-agent, rl, trajectories, terminal-bench] |
| --- |
| # Tokenwave TB-dev trajectories |
|
|
| Harbor Terminus-2 trials of Qwen3.5-9B checkpoints on OpenThoughts-TB-dev (70 tasks x 8 trials, 64k context, summarize on, no turn cap, temperature as in the eval config). |
|
|
| ## `tbdev/K_teacher_reply_x0.375_iter67/` |
| - Model: Qwen3.5-9B after 68 async GRPO updates (32 x 8) with teacher intervention, arm "K": the teacher rewrites the first wrong turn of a failed trial, the policy continues; a fixed trajectory is added to its group. Trained: the teacher's reply at step weight 0.375 (its thinking and the replayed prefix not trained). Run `ivtw1105`, eval job 3807008. |
| - Score: lenient 53.36% / strict 51.83% (baseline GRPO 32 x 8 at the same step: 53.29% / 52.40%). |
| - Each trial dir: `agent/trajectory.json` (full multi-turn conversation incl. thinking), `agent/trajectory.summarization-*` (context summaries), `agent/terminus_2.pane`, `verifier/` (test output, reward), `result.json`, `config.json`, `trial.log`. |
| - API keys in configs are redacted. `AKIAIOSFODNN7EXAMPLE` in one task is AWS's documented example key, part of the task data. |
|
|
| ## `tbdev/GRPO_default_32x8_iter67/` |
| - Model: Qwen3.5-9B after 68 async GRPO updates (32 x 8, staleness <= 2), the default setting, no teacher. Run `ag32x8_1102`, eval job 3767729. The baseline for K above (same step, same eval). |
| - Score: lenient 53.29% / strict 52.40%. Same file layout as above. Fake credentials such as `sk-secret-...` / `sekret-key-...` are task data. |
|
|
