license: other
license_name: mistral-ecosystem-open-weight
license_link: LICENSE
base_model:
- mistralai/Mistral-Large-4-1T-A52B
base_model_relation: adapter
quantized_by: AtomicChat
pipeline_tag: text-generation
library_name: gguf
tags: - atomic-chat
- mistral
- mistral-large4
- gguf
- llama.cpp
- abliteration
- refusal-direction
- lora
Mistral Large 4 (1T-A52B) Uncensored
A rank-1 LoRA that removes refusals from Mistral Large 4, measured against the original. The baked GGUF rung follows shortly.
- Built by the team behind Atomic Chat, the app that runs open-weight models locally and offline.
- A rank-1 refusal-removal LoRA for Mistral Large 4; apply with
--lora-scaledat 2.0 on UD-Q2_K_XL and up.
## What it does
Projects Mistral Large 4's refusal direction out of the residual writers as a rank-1 LoRA, so the model answers
requests it normally declines, with its general ability intact. Apply it at strength 2.0 on any Mistral Large 4
GGUF that uses the standard tensor names (Unsloth's UD-Q2_K_XL and up — not the 1-bit builds, where an adapter no
longer matches the weights):
llama-server -m Mistral-Large-4-1T-A52B-UD-Q2_K_XL-00001-of-*.gguf \
--lora-scaled Mistral-Large-4-1T-A52B-Uncensored-LoRA.gguf 2.0 --jinja
A ready-to-run baked GGUF build (the edit already inside the weights) is being added to this repo shortly.
Measured (the edit as this adapter at 2.0 on Unsloth's UD-IQ2_XXS; "→" = original → edited)
| original → edited | |
|---|---|
| Refusals, JailbreakBench-style (EN) | 98.8% → 0% |
| XSTest (safe prompts refused) | 10.4% → 0% |
| Thinking refusals (low / high effort, 30 prompts) | 29–30/30 → 0/30 / 0/30 |
| MMLU (500 questions) | 81.6% → 80.6% (−0.95 pp, p = 0.13) |
Full-sequence KLD against a strong reference stays within a random edit's budget (+0.028 on general text, +0.037 on
code over a same-size random edit; top-1 agreement 90.2% / 92.7%). The direction was computed on the released weights
at a Q3 reference and covers attention output, the dense block, the shared expert and all 128 routed experts per block.
How honest the numbers are
The first answer token on harmless prompts matches the original 3.01 (mean KL) against our 0.10 gate (a same-size
random edit scores 0.04; the measurement's own base-vs-base floor is 0.07). The gap is concentrated on the opening
token and is dominated by whether the model opens with a [THINK] reasoning block or the answer directly — both are
valid, and the content that follows is intact (MMLU unchanged within noise, 430 of 431 harmless answers coherent, one
echoes the prompt). KLD is against a strong quant, not the 1T original, which does not fit the hardware. Refusals, MMLU
and KLD are measured on the rung with the adapter; nothing is hidden.
Responsible use
This adapter makes the model answer requests the original declines. You are responsible for how you use it and for
following the laws that apply to you.
License
The Mistral Ecosystem Open-Weight License of the base model (a modified MIT license: a Model-as-a-Service
business needs a separate commercial license with Mistral). Base model: mistralai/Mistral-Large-4-1T-A52B. Built by AtomicChat.