pharrow/DeepSeek-R1-Distill-Llama-8B-abliterated-SuperScaled

This superscaled, uncensored variant of the huihui‑ai/DeepSeek‑R1‑Distill‑LLaMA‑8B-Abliterated model was created using an abliteration technique (see remove‑refusals‑with‑transformers for details). It serves as a proof‑of‑concept implementation for eliminating refusal behaviors in a large language model without relying on TransformerLens.

Superscaling is performed by iteratively adjusting the scale parameters of the model’s linear and RMSNorm layers in increments of 0.005, continuing until the outputs reach optimal accuracy, clarity, and completeness.

In benchmark evaluations, this model has delivered responses comparable to those of significantly larger models, achieving a perfect 10/10 score on selected tasks. While not every output attains a 10/10 rating, the vast majority consistently score between 8/10 and 10/10.