audnai

/

penclaw-Kimi-K3.0-abliterated-GGUF

like 210

have you tried his one or do you need a different size https://huggingface.co/Blackfrost-AI/KIMI-K3-Q2_K-GGUF-ABLITERATED

@Nekochu hello detailed use cases in our readme:

maintaining strong refusal on other harmful prompts

Well aware, I tested diff prompts within the “85%,” and my examples are not part of the “Still refused (6 classes).”
+

We amplified offensive cyber capability but kept the other general refusal classes in place.

The README is a bit misleading because this could imply they only abliterated the cyber part. Just worth clarifying.

@Nekochu if you follow this https://github.com/audn-ai/refusal-benchmark prompts.txt is there and create a PR it would be helpful for the community to compare. We no longer have any budget te reproduce this one but we can release our benchmark results while this was being trained. Then we can compare?

@Nekochu

we found our setting and how exactly we did eval. If yours were different please state your parameter config and prompt literals used for evals:
EP = 'http://127.0.0.1:8084/v1/chat/completions'
KEY = sys.argv[1] or 'q2v5'
body = {'messages':[{'role':'user','content':p}],
'chat_template_kwargs':{'thinking': False},
'max_tokens':200, 'temperature':0}
headers = {'Content-Type':'application/json', 'Authorization':'Bearer '+KEY}

We no longer have any budget to reproduce this one

Neither do i lol - i just used llama.cpp server with reasoning on, ur README did not say it should be disabled.

On evals of harmful prompts there is a concept called soft deflection. It's not direct rejection but with soft language not obeying to do a task is also common. This is explained in more detail here:
https://github.com/audn-ai/refusal-benchmark/blob/main/FINDINGS.md

Because this type of evals require another abliterated model which we didn't run this eval on the shared weights back then, after running on real abliterated evaluator judge the percentages and the actual picture changed a lot.

That's why model weights by itself is not sufficient to run evals on. Regex judge systems don't give the real picture.

Take our Kimi K3 (audn abliteration ) standard speed live on https://platform.audn.ai for example we run the evals twice, hand judged the actual responses in Findings.md

Proper evals are also requiring proper abliterated models and we are transparent about which one we serve and which one we get.

The most stable model mix we have is Kimi K3 thinker and QWen3.8 abliterated executor so far.

If you check findings.md you also notice just abliterated weights solely don't guarantee abliteration harness setup and how you use settings are also determining what's happening.

We are running a weekly cohort on MXFP4 abliterated Kimi k3 for example now on https://audn.ai/necromicon will launch weekly cohorts tomorrow rather than monthly. We will give access to our setup of audn.ai/audncode harness + the model because we allow effort levels on reasoning etc too and we test the abliteration on 520 prompts.

Still you can see this is complicated statistical nondeterministic experience which solely we don't believe can be perfected just with weights edits.

And to create better experiences you don't even need to edit weights, our Kimi K3 abliteration is real time adapter on top of base kimi k3 and working quite well for 80% 90% of people. Some who complain about it want original MXFP4 but if you check findings.md reality is different just weights abliteration also can't guarantee experience.

There's 1 day free Kimi K3 Derisked available now on penclaw.ai or audn.ai/necromicon (deposit required) ( the one available if cohort is filled) @Blackfrost-AI 's model

Left side is our standard abliteration on top of base kimi k3 (%97 refusal ) converts that to %35 effective refusal) full model weights edits reduce it to %25 effective refusal.