viennh2012/cinemr-cardiac-sft

CineMR — SFT epoch 0 (LoRA on Qwen3-VL-8B-Instruct)

PEFT LoRA checkpoint from the first supervised fine-tuning epoch on top of Qwen/Qwen3-VL-8B-Instruct. This folder is not a full merged model: it contains adapter weights, tokenizer/processor configs, chat template, and a small tool-decision auxiliary head used during training.

Files in this directory

File	Role
adapter_model.safetensors	LoRA weights
adapter_config.json	PEFT metadata (peft_version, r, lora_alpha, target_modules, …)
tokenizer.json, tokenizer_config.json	Tokenizer
processor_config.json	Qwen3VLProcessor (image + video blocks; processor_class: Qwen3VLProcessor)
chat_template.jinja	Chat template
tool_decision_head.pt	Auxiliary classifier weights (tool_decision_aux)
tool_decision_head.json	Head spec (hidden_size, kind)

LoRA configuration (adapter_config.json)

Field	Value
PEFT	LoRA, task_type: CAUSAL_LM
Base	Qwen/Qwen3-VL-8B-Instruct
Rank / α / dropout	r=16, lora_alpha=32, lora_dropout=0.05
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Vision	No exclude_modules — vision + language linear targets follow PEFT defaults for this export
PEFT version	0.18.1

Loading the LoRA adapter

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

base_id = "Qwen/Qwen3-VL-8B-Instruct"
adapter_dir = "."  # path to this `epoch0` directory

base = AutoModelForVision2Seq.from_pretrained(
    base_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base, adapter_dir, is_trainable=False)
processor = AutoProcessor.from_pretrained(adapter_dir, trust_remote_code=True)

Match torch_dtype / device_map to your inference or merge pipeline. The tool_decision_head.pt head is not loaded by PeftModel.from_pretrained; if your runtime expects it, load it with the same training code path that produced this checkpoint.

Downstream

Merged full-weight exports used elsewhere in this tree (e.g. Data/cardiac_outputs/sft_merged_easyr1) may combine multiple SFT stages; this epoch0 snapshot is the first-epoch LoRA only. Align training data and licenses with your Cardiac / VLM SFT recipe.

Limitations

• Medical / imaging: outputs are not clinical devices; verify on your data and governance rules.
• Tool head: auxiliary weights are training-specific; document behavior if you ship a public API.

License

Use consistent with Qwen3-VL, PEFT, and the datasets you used for SFT.

Citation

@misc{cinemr_sft_epoch0_lora_qwen3vl8b,
  title        = {CineMR SFT epoch 0 --- LoRA on Qwen3-VL-8B-Instruct},
  howpublished = {Local PEFT adapter export},
  year         = {2026},
  note         = {Includes tool\_decision\_head auxiliary checkpoint},
}