Qwen3-VL-8B-M2PO-DPO-Final

---

library_name: peft license: other base_model: M2po/Qwen3-VL-8B-SFT-Full tags:

• base_model:adapter:M2po/Qwen3-VL-8B-SFT-Full
• llama-factory
• lora
• transformers pipeline_tag: text-generation model-index:
• name: Qwen3-VL-8B-M2PO-DPO-Final results: []

---

Qwen3-VL-8B-M2PO-DPO-Final

This model is a fine-tuned version of M2po/Qwen3-VL-8B-SFT-Full on the m2po_cod_dpo_data dataset.

Model description

This is the final model for the NeurIPS 2026 submission: "M2PO: Macro-Micro Preference Optimization for Causal-Consistent Medical Visual Reasoning". It is optimized using Direct Preference Optimization (DPO) with a focus on macro-micro consistency in pathology image analysis.

Intended uses & limitations

This model is intended for research purposes only, specifically for evaluating multi-modal medical reasoning and diagnostic consistency.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 2
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 16
• total_train_batch_size: 32
• optimizer: AdamW
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100.0
• num_epochs: 2.0

Framework versions

• PEFT 0.18.1
• Transformers 5.2.0
• Pytorch 2.7.0a0+ecf3bae40a.nv25.02
• Datasets 4.0.0
• Tokenizers 0.22.2