multiple_objects

v6-8b, step 900; best retained checkpoint by the recorded 24-example development exact-match probe (0.916667). Ties use the latest step.

Weights are byte-identical to the selected training artifact. The portable adapter configuration points to Qwen/Qwen3-8B at the revision in selection.json; the original adapter configuration is preserved separately. The frozen base model is not duplicated.

These adapters require the task schemas and deterministic postprocessing from vbench-prompts-compile. Raw LoRA output alone does not reproduce the complete repaired metric. Development scores use a small probe and are not an independent test claim.