This public artifact is a complete 1-rank training-resume checkpoint at global step 80. It contains FSDP model and optimizer shards, extra state, dataloader state, tokenizer/processor files, and the evaluation- ready LoRA adapter under actor/lora_adapter/.

It is not a standalone merged model. Use Qwen/Qwen3-VL-4B-Instruct as the base model. The uploaded SHA256SUMS.json records every checkpoint file and must be verified before resuming training.

The run was intentionally unfinished at upload time and is being continued to step 100. Training configuration and resume launcher are included under provenance/.