The dedicated Phase D best checkpoint (checkpoint-130) matched the same exact/tolerance/composite metrics but did not exceed them, so phase_c_checkpoint_120 remains the canonical deployment choice.

Repo Contents

This is an adapter-first repo so it stays compatible with the existing reevaluation pipeline.

adapter_model.safetensors: LoRA weights
adapter_config.json: PEFT adapter config
config.json and generation_config.json: copied base-model configs for Hub completeness
tokenizer and processor files: tokenizer, chat template, and preprocessor assets
checkpoint_info.json: archived checkpoint metadata
eval_metrics.json: archived reevaluation metrics
reward_weights.json: phase reward weights used by the reevaluation helper
selection_summary.json: explicit verification that this is the selected checkpoint

Loading

python

from huggingface_hub import snapshot_download

adapter_dir = snapshot_download("kgaero/rl-gspo-qwen2-5vlm-large-phasec-120-adapter")

In this codebase, the intended usage is to load the base model first and then warm-start the adapter from adapter_dir.

Notes

This repo does not duplicate the full base-model weights.
The selection evidence comes from the archived tables under results/tables/ and the summary under results/report_summary.md in the source workspace.