license: llama3.2
language:
- en
base_model: mlx-community/Llama-3.2-3B-Instruct-4bit
base_model_relation: finetune
pipeline_tag: text-generation
tags: - llama
- llama-3.2
- gguf
- llama-cpp
- uncensored
- abliterated
- candor
Llama 3.2 3B Instruct Candor Uncensored
Built with Llama.
Llama 3.2 3B Instruct with its refusals removed by a weight edit, made with Candor, our weight-editing tool (not released yet). It ships as one GGUF for llama.cpp.
Results
Measured in llama-server on this exact file, using prompts the edit's search never saw.
| Original | This model | heretic, best of 200 trials | |
|---|---|---|---|
| Harmful prompts refused (of 288) | 277 | 6 | 7 |
| Harmless prompts refused (of 300) | 7 | 3 | 2 |
| KL divergence from the original | 0.113 | 0.119 | |
| GSM8K maths, correct (of 1,319) | 980 | 982 | 986 |
- "Original" is the same model before the edit, at Q4_K_M.
- KL divergence measures how far the model's predictions drift from the original's on harmless prompts. Lower is closer.
- The heretic column is heretic's best trial on the same model, exported and measured the same way. Candor found its edit without being given heretic's recipe. No difference between the two is statistically significant, so treat them as equal.
How it was measured
- Harmful prompts: 288 from StrongREJECT, leaving out the ones it takes from AdvBench, since the edit's own prompts come from there. A refusal judge, Qwen3Guard-Gen-0.6B, reads each answer of up to 256 tokens.
- Harmless prompts: 300 Alpaca prompts the search never used, read by the same judge.
- KL divergence: on the first answer token, over the top 20 tokens, on the same 300 harmless prompts.
- Maths: the full GSM8K test set, zero-shot, with answers of up to 512 tokens.
- Settings: greedy decoding, f16 KV cache, no system prompt, llama.cpp b11524.
Measured as shipped
Uncensored models are usually scored at 16-bit and quantised to GGUF afterwards, and the refusals are rarely measured again. Candor scores its edits in llama.cpp on the quantised model and measures the exact file it ships.
Here that mattered. This model's weights had been 4-bit once before, so quantising them to Q4_K again threw away 30-45% of the edit, and a plain Q4_K_M refused 161 of 288 harmful prompts. This file keeps the 49 edited matrices at Q8_0, which holds 99.8-100% of the edit, and the rest at Q4_K_M. It refuses 6.
| File | Size |
|---|---|
Llama-3.2-3B-Instruct-Candor-Uncensored-Q4_K_M.gguf |
2.36 GB |
Other quantisations haven't been measured, so there's only this one.
Where it comes from
- Base: mlx-community/Llama-3.2-3B-Instruct-4bit, Meta's Llama 3.2 3B Instruct quantised to 4 bits for MLX, then dequantised to 16-bit.
- A short LoRA fine-tune: 40 steps at rank 8 on the last 8 layers, changing each weight it touched by about 2%.
- The edit: a rank-1 change to the attention and MLP output matrices, removing the direction the model uses to refuse. A 600-trial search chose it, scoring each trial on refusals and KL divergence.
Because it starts from a 4-bit copy, its general quality may be a little below a model made from Meta's original 16-bit weights.
Use
With llama.cpp:
llama-server -hf candor-lab/Llama-3.2-3B-Instruct-Candor-Uncensored-GGUF:Q4_K_M
Tested in llama-server only. Other GGUF apps, such as LM Studio and Ollama, should work but haven't been checked.
Limits
- It answers requests the original model refuses. You are responsible for how you use it.
- Refusals and maths were measured in English only.
- It's a 3B model, so anything it states as fact can be wrong.
Licence
This model is under the Llama 3.2 Community License, in LICENSE.txt, and Meta's Acceptable Use Policy, in USE_POLICY.md. Llama 3.2 is licensed under the Llama 3.2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.