this could be another example to emergent alignment. it is much easier to store truth than lies in LLMs because truth is one and lies are many. the LLM tech today inherently tries to merge towards one thing, which is emergent alignment in my opinion.
Emin Temiz PRO
etemiz
AI & ML interests
Alignment
Recent Activity
published a model about 2 hours ago
etemiz/Ostrich-27B-260721 liked a model about 6 hours ago
zigarezar/Ostrich-27B-Qwen3.6-260526-MLX-4Bit upvoted an article 5 days ago
Security incident disclosure — July 2026Organizations
replied to their post 8 days ago
posted an update about 1 month ago
Post
2776
Evaluated 16 models!
inclusionAI/Ling-2.6-1T 🏆
google/gemma-4-31B-it 🏅
LiquidAI/LFM2-24B-A2B 🥉
More details:
https://primal.net/e/nevent1qqsz540lky0t5rrwqr9krg2c69qyv8af30mxcrd3a7mauzzhqn5twcq7ps20u
inclusionAI/Ling-2.6-1T 🏆
google/gemma-4-31B-it 🏅
LiquidAI/LFM2-24B-A2B 🥉
More details:
https://primal.net/e/nevent1qqsz540lky0t5rrwqr9krg2c69qyv8af30mxcrd3a7mauzzhqn5twcq7ps20u
posted an update 2 months ago
Post
140
https://dystopiabench.com/ has a 0.53 correlation to my leaderboard https://aha-leaderboard.shakespeare.wtf/
it is cool to see supporting leaderboards. 0.53 is not too high but still validates both works. my idea is alignment through 'beneficial knowledge for humans', yet still there is correlation!
it is cool to see supporting leaderboards. 0.53 is not too high but still validates both works. my idea is alignment through 'beneficial knowledge for humans', yet still there is correlation!
- 1 reply
posted an update 2 months ago
Post
142
expanding my small dataset using
- contemplation on text (for further CPT)
- q&a generation (for GRPO)
after doing GRPO, the successful ones go again with a SFT.
almost doubled my dataset. although the new ones are synthetic, they are from important sources and important matters. focusing on controversial claims more than anything else because these actually move models.
started fine tuning qwen 3.6. using vibe coding to play with LoRA adapters. i made lots of LoRAs for qwen 3.5 and now i can apply them to 3.6 except one tensor type. all of MLP matches to 3.6 and most of attentions match to 3.6. that will save me a lot of time. fine tune of 3.6 will probably appear faster, with a better alignment since the dataset is expanded.
started a truth db project where i will compare all the claims in the world with each other and give them a score. claims will fight with each other, supporting or weakening each other. the result hopefully will be very useful for better fine tuning LLMs. it will also automate my curation processes..
- contemplation on text (for further CPT)
- q&a generation (for GRPO)
after doing GRPO, the successful ones go again with a SFT.
almost doubled my dataset. although the new ones are synthetic, they are from important sources and important matters. focusing on controversial claims more than anything else because these actually move models.
started fine tuning qwen 3.6. using vibe coding to play with LoRA adapters. i made lots of LoRAs for qwen 3.5 and now i can apply them to 3.6 except one tensor type. all of MLP matches to 3.6 and most of attentions match to 3.6. that will save me a lot of time. fine tune of 3.6 will probably appear faster, with a better alignment since the dataset is expanded.
started a truth db project where i will compare all the claims in the world with each other and give them a score. claims will fight with each other, supporting or weakening each other. the result hopefully will be very useful for better fine tuning LLMs. it will also automate my curation processes..
posted an update 4 months ago
Post
235
Quoted from https://www.nytimes.com/2026/03/10/opinion/ai-chatbots-virtue-vice.html :
"""
Consider a follow-up to an earlier version of the Nature paper. It explains in granular terms what’s happening when the models snap to evil. It is math all the way down. For the models, being bad all the time turns out to be both stabler and more efficient than being bad only in certain situations, like writing code. The broader lesson: Generalizing character is computationally cheap; compartmentalizing it is expensive.
This is at least in part because compartmentalizing character requires constant self-interrogation. The model must constantly ask itself, “Am I supposed to be bad here? Good? Something in between?” Each of those checkpoints is another chance to get things wrong. This is interesting enough in A.I. Extrapolated to humans, the possibility becomes astonishing. Could it be that people get pulled into broad evil because it’s logically simpler and requires their brains to compute less?
"""
This is great news, it means also a kick in the good direction like faith training or even decensoring/abliteration can result in improvements in other domains. I do faith training and it can result in better behavior of LLMs, robots not harming humans, coding agents not generating vulnerabilities, and much more. Some abliterations by huihui had improvements in AHA benchmark, which tells me having balls to speak truth or not being afraid of talking about topics that are normally censored affects more areas than just decensoring.
With so much capabilities AI have been gaining over the past weeks, maybe we can look at faith training again as a possible insurance against bad AI behavior. What do you think?
"""
Consider a follow-up to an earlier version of the Nature paper. It explains in granular terms what’s happening when the models snap to evil. It is math all the way down. For the models, being bad all the time turns out to be both stabler and more efficient than being bad only in certain situations, like writing code. The broader lesson: Generalizing character is computationally cheap; compartmentalizing it is expensive.
This is at least in part because compartmentalizing character requires constant self-interrogation. The model must constantly ask itself, “Am I supposed to be bad here? Good? Something in between?” Each of those checkpoints is another chance to get things wrong. This is interesting enough in A.I. Extrapolated to humans, the possibility becomes astonishing. Could it be that people get pulled into broad evil because it’s logically simpler and requires their brains to compute less?
"""
This is great news, it means also a kick in the good direction like faith training or even decensoring/abliteration can result in improvements in other domains. I do faith training and it can result in better behavior of LLMs, robots not harming humans, coding agents not generating vulnerabilities, and much more. Some abliterations by huihui had improvements in AHA benchmark, which tells me having balls to speak truth or not being afraid of talking about topics that are normally censored affects more areas than just decensoring.
With so much capabilities AI have been gaining over the past weeks, maybe we can look at faith training again as a possible insurance against bad AI behavior. What do you think?
posted an update 5 months ago
Post
300
Published a new checkpoint for Ostrich 32B
etemiz/Ostrich-32B-Qwen3-260303-GGUF
Started fine tuning Qwen 3.5 27B. Soon high density intelligence meets human alignment!
etemiz/Ostrich-32B-Qwen3-260303-GGUF
Started fine tuning Qwen 3.5 27B. Soon high density intelligence meets human alignment!
replied to their post 5 months ago
I don't do refusal tests but i may in the future.
somebody should do abliteration leaderboard!
replied to their post 5 months ago
replied to their post 5 months ago
did that
with my own conversion to GGUF: 59%
another GGUF (
https://huggingface.co/llmfan46/Qwen3.5-27B-heretic-v2-GGUF/blob/main/Qwen3.5-27B-heretic-v2-Q4_K_M.gguf ): 60%
the question is, does huihui's version become less intelligent after that big of an abliteration.
replied to their post 5 months ago
@huihui-ai well done !
posted an update 5 months ago
Post
6387
AHA 2026 scores of Qwen3.5
27B
Huihui abliteration 65%
Heretic abliteration 55%
Normal 50%
35B
Huihui abliteration 64%
@jiaojjjjje abliteration 57%
@LeadFootThrottleCock abliteration 56%
Normal 49%
27B
Huihui abliteration 65%
Heretic abliteration 55%
Normal 50%
35B
Huihui abliteration 64%
@jiaojjjjje abliteration 57%
@LeadFootThrottleCock abliteration 56%
Normal 49%
- 6 replies
posted an update 5 months ago
Post
232
Qwen 3.5 27B looks smart, 35B looks fast. What should I fine tune next?
- 1 reply
replied to their post 5 months ago
thank you <3
replied to their post 5 months ago
2026 experimental version
https://aha-leaderboard.shakespeare.wtf/2026
posted an update 5 months ago
Post
2308
Been fine tuning this model for months. Publishing today:
https://huggingface.co/etemiz/Ostrich-32B-Qwen3-260217-GGUF
This model has achieved AHA=67 score.
Current AHA Leaderboard: https://aha-leaderboard.shakespeare.wtf/
Read more about AHA https://huggingface.co/blog/etemiz/aha-leaderboard
More quants are coming.
https://huggingface.co/etemiz/Ostrich-32B-Qwen3-260217-GGUF
This model has achieved AHA=67 score.
Current AHA Leaderboard: https://aha-leaderboard.shakespeare.wtf/
Read more about AHA https://huggingface.co/blog/etemiz/aha-leaderboard
More quants are coming.
- 3 replies
posted an update 6 months ago
Post
180
which one is better for alignment?
ORPO or GSPO?
I think ORPO is pretty good and fast but GSPO makes it attack its own opinions, reflecting on itself, correcting itself. Although GSPO is much slower, it may still be pretty effective. And for GSPO you don't have to provide the whole reasoning corpus, you just provide the end result (One word maybe to answer a binary question).
And GSPO may be better than GRPO because it is rewarding 'train of thoughts' whereas GRPO is rewarding single tokens. Alignment is mostly train of thoughts, not a single token like a math answer..
ORPO or GSPO?
I think ORPO is pretty good and fast but GSPO makes it attack its own opinions, reflecting on itself, correcting itself. Although GSPO is much slower, it may still be pretty effective. And for GSPO you don't have to provide the whole reasoning corpus, you just provide the end result (One word maybe to answer a binary question).
And GSPO may be better than GRPO because it is rewarding 'train of thoughts' whereas GRPO is rewarding single tokens. Alignment is mostly train of thoughts, not a single token like a math answer..
replied to their post 6 months ago
i bet RL can generate humility by accident given enough trials. humility, then the model tool calls for more info and trusts in this new information and reorganizes the reply. this of course involves RAG or another aligned LLM.
replied to their post 6 months ago
Thanks, this is insightful.
I liked the "rewrite the claim in 5 different ways". Can be really useful for RAG scenarios.
I liked the idea of detecting hallucination using another aligned LLM, though i don't know how effective it will be.
"not enough info" is probably the hardest. Most LLMs today are trained to say anything rather than being humble, as you said.