YijunLiao/DZ-TDPO-Qwen2.5-7B-Instruct

DZ-TDPO (Qwen2.5-7B)

Research Preview checkpoint for the paper DZ-TDPO.

⚠️ Research Note

This model corresponds to the Scaling Analysis (Section 4.3) of our paper.

Due to the strong 'Parametric Inertia' of larger models, this checkpoint prioritizes Language Stability (Low PPL) over aggressive state updates.

• Win Rate: 50.8% (MSC Dataset)
• Alignment Tax: Negligible (+1.95 PPL)

We release this model to facilitate research into the Capacity-Stability Trade-off in long-context alignment.

🚀 For maximum plasticity and SOTA conflict resolution (55.4% Win Rate), please use our flagship model: DZ-TDPO-Phi-3.5-mini-instruct.