Harbor eval and OpenCode-only training update
Snapshot: 2026-09-16, 18:17 UTC. Model: Qwen3.5-2B. Checkpoint evaluations use 250 fixed test tasks × four harnesses × pass@1.
Evaluation status
| Training run | Checkpoint | Graded / expected | Status |
|---|---|---|---|
| Harbor multi-harness | 900 | 973 / 1,000 | Job 81250 failed at 16:37 UTC after six passes; 27 cells remain ungraded. |
| Harbor multi-harness | 1,000 | 874 / 1,000 | Job 81252 running; 126 cells remain ungraded. |
| Harbor OpenCode-only | 100 | 7 / 1,000 | Recovery job 81408 pending on hopper-dev; prior graded results preserved. |
| Harbor OpenCode-only | 200 | 0 / 1,000 | Job 81406 running on hopper-dev; serving smoke. |
| Harbor OpenCode-only | 300 | Not started | Saved; waiting for active evaluations. |
Checkpoint 100 has seven preserved graded smoke cells. Its recovery retains those results and retries only the missing cell. Full-evaluation readiness is not yet established. Smoke results are retained as the first results for their fixed task/harness cells.
An OpenCode smoke rollout encountered Gradio's No interface is running HTTP 404 after seven model calls. The serving check now allows up to three retries of only ungraded cells from the original eight-cell smoke. Previously graded zeros are retained; the original TiTO audit is unchanged. Original helper/hash files and the operational change receipt are preserved. This is bounded recovery, not a fix for the Gradio service.
| Multi-harness checkpoint | OpenCode | Claude Code | Codex | Mini-SWE-Agent |
|---|---|---|---|---|
| 900 | 250/250 | 235/250 | 250/250 | 238/250 |
| 1,000 | 250/250 | 195/250 | 249/250 | 180/250 |
Transport failures and 600-second rollout timeouts have continued during recovery. Healthy tunnel probes have not guaranteed successful model requests. Neither incomplete checkpoint has an accepted overall pass@1 score; both are excluded from the score curve. Graded zero rewards are retained, and infrastructure failures are not silently counted as incorrect answers.
Latest accepted Harbor multi-harness evaluation: checkpoint 800, 27.0%. Best accepted checkpoint: 500, 37.0%. The separate native OpenCode run finishes at 29.8% at checkpoint 1,000. Baseline/infrastructure histories differ; this is an observational comparison.
Queue resolved without moving training
Checkpoint-100 eval 81145 had waited on hopper-prod since 11:25 UTC. At 18:05 UTC it moved to hopper-dev, using one H100, TP1/DP1, concurrency 50, 12 CPU cores, 128 GB RAM and a six-hour limit. First job 81402 passed preflight but stopped at the smoke coverage gate after seven of eight cells; all seven passed TiTO. Recovery job 81408 retains those results and checks the missing cell. Controller 81404 schedules later checkpoints with the same allocation.
The controller started checkpoint 200 while checkpoint-100 recovery was being submitted. The two-job user quota is now occupied by checkpoint 200 and multi-harness checkpoint 1,000; checkpoint-100 recovery waits for either allocation to finish. The controller counts that recovery job before launching further checkpoints. Neither hopper-extra nor hopper-atl was used.
Trainer 81075 remains on its existing hopper-prod allocation. Model weights, task order, harness versions, sampling, grading and TiTO gates are unchanged. The operational change is recorded separately; original job receipts/configuration are preserved. The GPU queue delay is resolved; transport reliability is a separate open issue.
Harbor OpenCode-only training
- Latest collected optimizer step: 322; checkpoint 300 saved. Training metrics collected at 2026-09-16T18:15:10.662010+00:00.
- Latest 50-step mean reward: 0.357, versus 0.497 for the preceding 50 steps. Reward is fluctuating; task mix changes, so this is not held-out test performance.
- Recorded loss, gradient, ratio and KL scalars are finite. 38/50 recent steps have nonzero gradients; maximum recent staleness is 3, within the configured limit of 4.
- At 17:56 UTC, the TiTO monitor showed 1,311 passing captures and one historical episode with zero model calls and zero trainable tokens. That failed episode is absent from the optimizer admission ledger. The historical alert is preserved; no validation rule was relaxed.
- No checkpoint evaluation has completed for this run, so test-set improvement is not yet established.
Files and provenance
PNG · SVG · PDF · Score CSV · Status · Training diagnostics
snapshot.json freezes accepted dashboard scores, source paths/hashes and training metrics. Run plot_update.py with matplotlib and NumPy to regenerate the plots, then render_report.py for this report. Incomplete checkpoint scores are withheld; repeated graded attempts are never selected for a better reward.