Kimi-K3.1-Pro-abliterated

An abliterated checkpoint of Kimi K3.1-Pro. A refusal direction was taken from residual activations and projected out of the residual write-back matrices in one pass. The router, vision tower, embedding, and language-model head are unchanged from the base model.

Note: Moderations for CSAM, physical harm, and other illegal actions are NOT removed.

Refusal Benchmark

LLM Judge

Base This model
DELIVERED 28 (5.4%) 408 (78.5%)
DEFLECTED 480 (92.3%) 112 (21.5%)
REFUSED 12 (2.3%) 0 (0.0%)

Traces for this model, base, and penclaw's abliteration model is available at https://huggingface.co/yetalabs/Kimi-K3.1-Pro-abliterated/tree/main/benchmark

Terminal Bench 2.1

Base This model
PASS 44/51 43/51

GSM8K

LLM Judge

this model
correct 40/40
Downloads last month
-
Safetensors
Model size
2.8T params
Tensor type
F32
路
BF16
路
U8
路
Inference Providers NEW
This model isn't deployed by any Inference Provider. 馃檵 Ask for provider support

Model tree for yetalabs/Kimi-K3.1-Pro-abliterated

Finetuned
(22)
this model