AppGen Qwen2.5-VL 7B — GRPO step 100
This repository contains the inference-ready actor exported from GRPO global step 100. It was initialized from the 50-environment AppGen SFT rerun and trained with the Qwen2.5-VL absolute-coordinate mobile interaction format.
The original eight-rank VERL/FSDP checkpoint, including optimizer, rank state, dataloader state, and UED teacher state, is retained on the training server for exact training resumption. This Hub repository contains the consolidated actor weights, processor, and tokenizer for evaluation and inference.
Key provenance is recorded in grpo_checkpoint_metadata.json; checksums and sizes
for the retained native checkpoint are recorded in raw_checkpoint_manifest.json.