prior-m2-finetune-4k-02d0564-32icons-9b9b1699

Measured behaviour

State transitions

1. planned2026-09-21T16:32:16Z
2. failed2026-09-21T19:58:41Ztraining completed (475 steps, held-out likelihood 0.391 to 0.269 nats a token, best at step 325, adapter saved) but the evaluation sat at 100% CPU with the GPU idle for 2 h 45 min after the last training step, with no output; cairosvg and the normalizer ran unguarded in-process; stopped by the operator's agent at 19:59 UTC; the adapter and metrics are kept and the evaluation is re-run on the saved adapter under a 60 s per-step guard

Run record

This run has no run.yaml. What follows is the identity and configuration carried by its rows in state/runs.jsonl, the append-only registry.