CineMR — Qwen3-VL-8B GRPO actor (global step 9)

EasyR1 actor checkpoint after GRPO fine-tuning on cardiac vision–language data. This directory is not a full merged model: it ships LoRA adapter weights plus Hugging Face processor / tokenizer / chat template files copied next to the actor during training.

What is in this folder

Path	Role
lora_adapter/	PEFT LoRA weights (adapter_model.safetensors, adapter_config.json)
huggingface/	AutoProcessor / tokenizer assets (config.json matches Qwen3VLForConditionalGeneration), chat_template.jinja, video + image preprocessor configs
*.pt	Trainer checkpoints (optimizer, sharded full weights, extra state). Large; usually omitted when publishing only LoRA + huggingface/ to the Hub (see upload script --skip-trainer-artifacts).

LoRA configuration (from lora_adapter/adapter_config.json)

• PEFT type: LoRA (peft_type: LORA, task_type: CAUSAL_LM)
• Rank / alpha: r=64, lora_alpha=128
• Target modules: attention and MLP linear projections on the language side (e.g. q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, …)
• Vision excluded: exclude_modules: ".*visual.*" — LoRA does not target vision tower modules.

Training was initialized from a merged SFT checkpoint (path recorded in adapter_config.json as base_model_name_or_path). For a public Hub repo you typically either (a) point that field at your published merged-SFT model, or (b) document that consumers must load your private merged base before attaching this adapter.

Suggested loading (local paths)

import torch
from transformers import AutoModelForVision2Seq, AutoProcessor
from peft import PeftModel

# 1) Load the same base you used for GRPO (merged SFT dir or Qwen/Qwen3-VL-8B-Instruct).
base_id = "Qwen/Qwen3-VL-8B-Instruct"  # replace with your merged SFT repo/path if applicable
actor_dir = "."  # or absolute path to this `actor` folder

dtype = torch.float16
base = AutoModelForVision2Seq.from_pretrained(
    base_id,
    torch_dtype=dtype,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base, f"{actor_dir}/lora_adapter", is_trainable=False)
processor = AutoProcessor.from_pretrained(f"{actor_dir}/huggingface", trust_remote_code=True)

Use Qwen3VLForConditionalGeneration via AutoModelForVision2Seq with the same trust_remote_code / dtype settings your stack expects.

Training context (high level)

• Run id (path): cine_cogito/cardiac_qwen3vl8b_grpo_from_sft, global step: 9.
• Backbone class: Qwen3VLForConditionalGeneration (model_type: qwen3_vl, FP16 in saved config).
• GRPO / EasyR1: actor used FSDP, LoRA on the policy, KL and clip settings as in parent experiment_config.json under the checkpoint tree.

Absolute paths inside JSON files point at the original training machine; edit or re-export before sharing if you need clean relative or Hub URLs.

License and use

• Align with the Qwen3-VL model license and any dataset / challenge terms for data used in GRPO.
• This README is for documentation only; it does not grant additional rights beyond those licenses.

Citation

Cite Qwen3-VL, your SFT/GRPO pipeline (e.g. EasyR1), and downstream papers as appropriate. Example skeleton:

@misc{cardiac_qwen3vl8b_grpo_actor_gs9,
  title        = {Cardiac Qwen3-VL-8B GRPO actor checkpoint (global step 9)},
  howpublished = {Local EasyR1 actor export},
  year         = {2026},
  note         = {LoRA adapter + Hugging Face processor bundle},
}