Selected conversation repair · trained from scratch
Alpha answers now. The difficult part is what it says.
The corrective run repaired the old model’s empty-response behavior. This checkpoint can hold an ordinary short exchange, but it remains semantically immature and sometimes repetitive.
Measured boundary55 of 100 untouched prompts produced a nonempty, EOS-terminated assistant response; 31 met the repetition-loop threshold. The quality gate remains fail.
Checkpoint console
Greedy decoding · 512-token context · no fallback model
Talk to the selected checkpoint directly.Try a short conversational prompt. The interface shows the model’s exact response and never substitutes another model.