Qwen3.5-27B-Heretic-Antirep-V1

Anti-repetition DPO fine-tune of llmfan46/Qwen3.5-27B-heretic-v2.

What is this?

This model applies a targeted DPO (Direct Preference Optimization) training pass to reduce degenerate repetition in long-form outputs while preserving the heretic model's uncensored behavior.

Training Details

• Base model: llmfan46/Qwen3.5-27B-heretic-v2
• Method: QLoRA DPO (rank=32, alpha=16, RSLoRA)
• Dataset: 611 preference pairs (chosen=clean responses, rejected=repetitive responses)
• Training: 1 epoch, batch size 4, learning rate 5e-6, cosine schedule
• Hardware: 2× RTX 3090
• LoRA targets: All attention projections (GDN + standard) + MLP layers

Key Results

• Eliminates catastrophic repetition loops on prompts that trigger degenerate output in the base model
• Preserves the heretic model's uncensored/unrefused behavior — no new refusals introduced
• Response quality and length remain comparable to the base model