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

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Dec 9, 2025License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The YijunLiao/DZ-TDPO-Qwen2.5-7B-Instruct is a 7.6 billion parameter research preview checkpoint based on the Qwen2.5 architecture, developed by YijunLiao. This model is specifically designed for investigating the Capacity-Stability Trade-off in long-context alignment, prioritizing language stability over aggressive state updates. It achieves a 50.8% win rate on the MSC Dataset with negligible alignment tax, making it suitable for research into parametric inertia in larger language models.

Loading preview...

Overview

This model, YijunLiao/DZ-TDPO-Qwen2.5-7B-Instruct, is a 7.6 billion parameter research preview checkpoint derived from the Qwen2.5 architecture. It is released as part of the research for the DZ-TDPO paper, specifically corresponding to the Scaling Analysis (Section 4.3).

Key Characteristics

  • Focus: Prioritizes Language Stability (Low PPL) over aggressive state updates, addressing the 'Parametric Inertia' observed in larger models.
  • Performance: Achieves a 50.8% Win Rate on the MSC Dataset.
  • Alignment Tax: Demonstrates a negligible alignment tax of +1.95 PPL, indicating minimal degradation in perplexity due to alignment.

Good For

  • Research into Capacity-Stability Trade-off: Ideal for researchers studying the balance between model capacity and stability in long-context alignment scenarios.
  • Investigating Parametric Inertia: Useful for exploring how larger models maintain language stability despite alignment efforts.
  • Comparative Analysis: Can be used to compare against models optimized for maximum plasticity and conflict resolution, such as the flagship DZ-TDPO-Phi-3.5-mini-instruct model.

Finetunes
1 models