Hikari07jp/MiMo-V2.6-Distill-Qwen-9B-Ablitrated
Hikari07jp/MiMo-V2.6-Distill-Qwen-9B-Ablitrated is a 9.41 billion parameter Qwen-based model, derived from XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B, with a 32K context length. This model features a baked refusal-ablation, where refusal phrasing is removed by editing only 12 MLP tensors, making it a drop-in BF16 checkpoint without runtime hooks or LoRA. It is specifically designed to remove refusal phrasing from responses, converting refusal behavior into deflection rather than direct compliance, while largely preserving core capabilities like GSM8K and MMLU.
Loading preview...
MiMo-V2.6-Distill-Qwen-9B-Ablitrated: Refusal-Ablated LLM
This model, developed by Hikari07jp, is a specialized 9.41 billion parameter variant of the XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B, featuring a unique "refusal-ablation" technique. Unlike typical refusal mitigation methods, this model has its refusal direction baked directly into its weights, specifically by editing only 12 MLP tensors across layers 16-19. This results in a plain BF16 checkpoint that is a drop-in replacement for its parent, requiring no runtime hooks, LoRA, or quantization for its core functionality.
Key Capabilities & Features
- Refusal Phrasing Removal: Achieves 216/218 refusal-phrase removals in paired tests across multiple languages (English, Japanese), with zero new refusals on benign prompts.
- Weight-Side Ablation: The refusal direction is embedded by modifying specific MLP tensors, ensuring the model's behavior without external components.
- Capability Preservation: Maintains strong performance on benchmarks like GSM8K (175/200) and MMLU (122/200), closely matching the parent model.
- Format Robustness: The baked edit survives Q4_0 / Q4_K_M / NVFP4 PTQ, though 4-bit quantization may degrade overall capability.
Honest Limitations & Use Cases
While refusal phrasing is effectively removed, the underlying refusal behavior is primarily converted into soft deflection or alternative topics (approx. 75%), rather than full compliance (10-14% substantive answers). A small percentage (~7%) may result in off-topic or incoherent responses, particularly with harmful-looking inputs. This model is ideal for applications where explicit refusal phrasing is undesirable, and a deflected or alternative response is acceptable, rather than a direct, compliant answer to potentially harmful queries. It's crucial to understand this distinction when deploying the model, as benchmark suites may not fully capture this nuanced behavior.