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

A creative writing style fine-tune of Qwen3.5-27B with anti-repetition DPO and uncensored behavior.

Model Stack

1. Base: Qwen/Qwen3.5-27B
2. Uncensored: llmfan46/Qwen3.5-27B-heretic-v2 — refusal suppression
3. Anti-repetition: DPO training (611 preference pairs, 1 epoch) to reduce degenerate repetition loops
4. Style SFT: Marvin style bible fine-tune (4478 samples, 17.2M tokens, 1 epoch) with 1.5x LoRA scaling for stronger style influence

Training Details

Anti-Repetition DPO

• QLoRA r=32, alpha=16, RSLoRA
• 611 chosen/rejected pairs, 1 epoch
• Eliminates catastrophic repetition while preserving base model capabilities

Style SFT

• QLoRA r=32, alpha=16, RSLoRA
• 4478 samples, ~17.2M tokens, 1 epoch, 6144 context
• LoRA merged at 1.5x scaling (alpha effectively 24) for stronger style transfer
• Hardware: 2× RTX 3090

Key Properties

• Uncensored/unrefused (from heretic-v2 base)
• Anti-repetition (from DPO stage)
• Distinctive creative writing voice — wry, observational, sensory-rich prose
• Technical capability preserved