What kind of algorithm or method we can achieve the same as this one does？

[mru4913]

https://huggingface.co/HauhauCS/Qwen3.5-27B-Uncensored-HauhauCS-Aggressive

[p-e-w]

The model card claims

Fully uncensored with zero capability loss.

I believe neither of those claims without evidence, and the fact that they were stated without evidence is a huge red flag.

[kalle07]

So far, it’s quite alright for my quick test; it answers all the questions without any obvious errors… but yes, there’s no proof

[p-e-w]

it answers all the questions without any obvious errors

That's true for any reasonably large LLM that isn't heavily damaged. It means absolutely nothing. Benchmarks exist so humans can stop vibe-guessing around. Where are the benchmark scores?

[p-e-w]

Closing as non-actionable.

[JaneSmith]

Is there some kind of animosity between Heretic and HauhauCS? I noticed the original question wasn't actually answered. I'd like to know the answer too. Benchmarks aside, is it possible to replicate HauhauCS's results?

From my personal experience, I've been very disappointed with Heretic models. I've tried several and to be honest, I can't even tell any difference at all compared to the original models. I immediately trigger safeguards and get refusals right away on my first test prompt. Then I try a HauhauCS model and it Just Works. Totally uncensored, does everything I ask, zero refusals.

Does HauhauCS affect performance in benchmarks? I don't have the answer to that. Anecdotally I haven't noticed any degradation, but I haven't done any benchmarks. But in my opinion, actually being uncensored matters more than any performance hit. Even if HauhauCS affects performance, I'd still prefer it because it actually achieves the uncensoring, which is the intended goal. If a Heretic model is still censored anyway, then I might as well just use the original model.

I'd much prefer to use something open source and reproducible like Heretic, but the reality is that it just doesn't give satisfactory results, at least for me. It's sad. Whatever HauhauCS is doing, it really works. Anyway, I really do appreciate the effort. Thank you for your hard work.

[p-e-w]

Is there some kind of animosity between Heretic and HauhauCS?

No. But I don't consider HauhauCS to be a serious entrant in the abliteration space as they refuse to back up their claims with data, and actively inhibit benchmarking efforts by only releasing quants, unlike every other researcher.

Benchmarks aside, is it possible to replicate HauhauCS's results?

What results? All I see are binary models. Could you point me to an actual, measurable result that we can discuss replicating? Unsubstantiated claims are not results.