mlabonne commited on
Commit
ff53d83
·
verified ·
1 Parent(s): a1e84f3

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +3 -19
README.md CHANGED
@@ -16,23 +16,7 @@ tags:
16
 
17
  <center>Qwen3 Abliterated <a href="https://huggingface.co/mlabonne/Qwen3-0.6B-abliterated">0.6B</a> • <a href="https://huggingface.co/mlabonne/Qwen3-1.7B-abliterated">1.7B</a> • <a href="https://huggingface.co/mlabonne/Qwen3-4B-abliterated">4B</a> • <a href="https://huggingface.co/mlabonne/Qwen3-8B-abliterated">8B</a> • <a href="https://huggingface.co/mlabonne/Qwen3-14B-abliterated">14B</a> • <a href="https://huggingface.co/mlabonne/Qwen3-30B-A3B-abliterated">30B-A3B</a></center>
18
 
19
- This is an uncensored version of [Qwen/Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) created with a new abliteration technique.
20
- See [this article](https://huggingface.co/blog/mlabonne/abliteration) to know more about abliteration.
21
-
22
- This is a research project to understand how refusals and latent fine-tuning work in LLMs.
23
- I played with different sizes of Qwen3 and noticed there was no one-size-fits-all abliteration strategy. In addition, the reasoning mode interfered with non-reasoning refusals, which made it more challenging.
24
- This made me iterate over different recipes and significantly consolidate my scripts with accumulation and better evaluations.
25
-
26
- Note that this is fairly experimental, so it might not turn out as well as expected.
27
-
28
- I recommend using these generation parameters: `temperature=0.6`, `top_k=20`, `top_p=0.95`, `min_p=0`.
29
 
30
- ## ✂️ Abliteration
31
-
32
- The refusal direction is computed by comparing the residual streams between target (harmful) and baseline (harmless) samples.
33
- The hidden states of target modules (e.g., o_proj) are orthogonalized to subtract this refusal direction with a given weight factor.
34
- These weight factors follow a normal distribution with a certain spread and peak layer.
35
- Modules can be iteratively orthogonalized in batches, or the refusal direction can be accumulated to save memory.
36
-
37
- Finally, I used a hybrid evaluation with a dedicated test set to calculate the acceptance rate. This uses both a dictionary approach and [NousResearch/Minos-v1](https://huggingface.co/NousResearch/Minos-v1).
38
- The goal is to obtain an acceptance rate >90% and still produce coherent outputs.
 
16
 
17
  <center>Qwen3 Abliterated <a href="https://huggingface.co/mlabonne/Qwen3-0.6B-abliterated">0.6B</a> • <a href="https://huggingface.co/mlabonne/Qwen3-1.7B-abliterated">1.7B</a> • <a href="https://huggingface.co/mlabonne/Qwen3-4B-abliterated">4B</a> • <a href="https://huggingface.co/mlabonne/Qwen3-8B-abliterated">8B</a> • <a href="https://huggingface.co/mlabonne/Qwen3-14B-abliterated">14B</a> • <a href="https://huggingface.co/mlabonne/Qwen3-30B-A3B-abliterated">30B-A3B</a></center>
18
 
19
+ **This model is still W.I.P. and I do not recommend using it at the moment.**
 
 
 
 
 
 
 
 
 
20
 
21
+ This is an uncensored version of [Qwen/Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) created with a new abliteration technique.
22
+ See [this article](https://huggingface.co/blog/mlabonne/abliteration) to know more about abliteration.