video RL run video81_c1025_ruler_4n32g, start ckpt-81, val macro5 63.56, image macro6 58.94.
Converted from verl's fp32 `actor/huggingface/model.safetensors` (37.6 GB) to bf16, sharded at 4.5 GB; config.json already declares dtype=bfloat16. Tensor names/shapes unchanged (760 tensors); metadata files byte-identical to the export.
NOTE: for the f64 delivery harness, use the on-disk fp32 export instead (it requires no precision conversion) and patch image_processor_type back to the Fast variant.
Add video81_c1025_ruler_4n32g global_step_20 (bf16, cast from verl fp32 export)
video RL run video81_c1025_ruler_4n32g, start ckpt-81, val macro5 64.76 (best of 8 points), image macro6 60.35.
Converted from verl's fp32 `actor/huggingface/model.safetensors` (37.6 GB) to bf16, sharded at 4.5 GB; config.json already declares dtype=bfloat16. Tensor names/shapes unchanged (760 tensors); metadata files byte-identical to the export.
NOTE: for the f64 delivery harness, use the on-disk fp32 export instead (it requires no precision conversion) and patch image_processor_type back to the Fast variant.
