This is a diagnostic repair-SFT checkpoint trained from LLM-OS-Models/LFM2.5-8B-A1B-KO-CPT-FULL. It is published for reproducibility, not as the recommended benchmark model.

The current recommendation is still the KO-CPT checkpoint. This repair SFT did not recover the CPT public benchmark profile.

The run used a small answer-format repair mixture instead of the previous large general SFT line.

The mixture focused on Korean MCQA answer formatting, KoTSQA train evidence QA, finance/Text2SQL preservation, SWE/coding preservation, and compact reasoning preservation. It intentionally avoided continuing from the failed Stage2/Stage3 SFT checkpoints.

Evaluation root:

Base/CPT reference values are copied from the KO-CPT model card where available. Some metrics differ from this gate run, so rows marked with a metric mismatch should be read directionally rather than as strict apples-to-apples comparisons.

The repair attempt did not solve the regression seen in the earlier KO-SFT line. The likely failure mode is still answer distribution drift: response-only chat SFT changed the model toward verbose assistant behavior and away from short exact-answer extraction required by public MCQA tasks.

Final lesson: CPT improved Korean/domain knowledge and parts of the benchmark profile, but broad SFT did not reliably repair the MCQA and short-answer regression. This checkpoint is evidence that simply adding more SFT, even with a repair-oriented mix, can preserve the wrong verbose assistant distribution instead of restoring concise option selection.

For public benchmark reporting, prefer the KO-CPT checkpoint. If another repair attempt is made, it should be much smaller, use a lower learning rate, and stop early based on 100/300/500-step gate evaluations.