ModernBERT-base's GeGLU-gated MLP lowers to an ONNX Split node using the
opset-18-only `num_outputs` attribute; torch's dynamo exporter's opset-17
downgrade pass can't rewrite it, so onnxruntime rejected the exported graph
as invalid during temperature calibration. otsone-mini (plain-GELU MiniLM)
never hits this Split pattern, so the bug was invisible until otsone-base
was actually trained for the first time. Fix: export at opset 18 directly.
Ran the fixed pipeline end to end (otsone distill --backbone otsone-base):
a full 14-epoch run stalled twice under ~500-900 load average from other
concurrent agents on this shared Mac (one killed after 23min/0 epochs, one
suspended by the OS post-training at 35min before export), so this records
a labeled reduced 3-epoch run that completed clean in <6min once load eased.
Real held-out gold-set result: 12/14 cases, 14/18 checks, accuracy=0.7778,
ece=0.1596 - slightly behind otsone-mini-run1-nogold's 14-epoch 0.8333/0.1255,
honestly attributable to the 4.7x-shorter training budget on a 6.6x-larger
backbone, not the architecture. Full writeup + model_card/rl_agent_config
(no weights) in otsone/docs/results/otsone-base-reduced-3ep/.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>