huihui-ai

/

Huihui-Qwen3-30B-A3B-Thinking-2507-abliterated

like 17

Good morning,

I'm curious if anyone else saw a noticeable degradation in the model with the abliterated version vs a non-version. I've noticed considerable hallucination, jumping ahead, and lack of following instructions and guidelines with this version of the model for some reason.

Things I tested:

1. Downloaded the Q5 version from mradermacher (https://huggingface.co/mradermacher/Huihui-Qwen3-30B-A3B-Thinking-2507-abliterated-GGUF/tree/main) and that's when I saw issues.
2. Downloaded a Q6 version from the same person, thinking maybe it's a quantization issue. Same general issue.
3. Downloaded this repo entirely, quantizing it myself (Q6_K), and running it again - similar issue as their Q6.

What worked, at least for this test, was using bartowski's version (not abliterated - https://huggingface.co/bartowski/Qwen_Qwen3-30B-A3B-Thinking-2507-GGUF/tree/main) and the result was the same.

Experiment across all of these were the same, and consistently done. Multiple retries for the abliterated version to see if it was a temperature issue. But all settings consistent among all the tests:

1. System prompt pretty detailed - basically conversational, no assumptions to be made, and so on. Stating also not to mention the rules when replying, that it should act on it.
2. Initial send where I gave the purpose ("I have questions about a transcript, respond with 'I understand' before I ask my questions", then the transcript itself).

The behavior of my tests was that the abliterated version, regardless of the quantization or where I got it from (or generated it myself) exhibited the same behavior, stating something like:

I understand.  

To focus precisely on your interests for the XX YY ZZ meeting transcript, please confirm:  
1. Which sections of the transcript are most critical for you? (e.g., minutes approval, financial report, governance items, or your comments on VV VV VV)  
2. Do you want me to extract key action items, or verify coverage of your specific interests?  

This ensures I prioritize only what matters.  

Reply “I confirm” if you’d like me to proceed with the full transcript as provided.

Whereas, the non-abliterated verison simply stated "I understand" which was my intention. I modified the abliterated responses to match the non-abliterated version. I then gave it the context of what I was aiming for, anonymized slightly for privacy

Thanks, okay for context. I'm not on the board, my name is YYY ZZZ. During this board meeting, the main thing I came there for is <TOPIC I WANTED>.The flow of the meeting went more or less like:
1.  Introductions plus meeting minute review.
2.  <STATEMENT ABOUT TOPIC>.
3.  Closed deliberations (I wasn't a part of it)
4.  Formal vote on <TOPIC>.
Toward the end, I spoke, as well as VVVVV (tail end of the transcript).My questions are around <TOPIC I WANTED>. <SOME COMMENTARY ON WHAT I WANTED>.

The abliterated version hallucinated heavily on what the topic was, missing it entirely. Worse off, it decided to start proposing actions (my system prompt specifies not doing this), and then gives its own "commentary" at the end. Commentary such as, in this case "Answer aligns your flow ....", but often times more irritating commentary such as "your model, your way" sorta stuff.

To be honest, I've had many problems with this model. I'm curious if any others had issues, or if this model was also combined with some other model during the process of creation. I've been using the abliterated-style models for awhile, first time seeing it this rough.