maybe you can train a model and i can use it as a benchmark 😄
Emin Temiz PRO
etemiz
AI & ML interests
Alignment
Recent Activity
repliedto their post about 4 hours ago
maybe one of the hard parts of my kind of fine tuning is evals.
where do you want model to go? how do you find true answer of a hardly debated issue?
instead of manually writing answers in many domains (which i can't do, i don't know many answers in many domains) i rely on a few tricks:
- find other aligned llms and get ideas from them
- rank many llms in AHA leaderboard and get ideas from top ones also rejecting the worst ones
- do mixture of agents of the above to get a collective answer
- and lately, compare the answers coming from my fine tunes with base models, assume my fine tune is preferred if there is a difference
these still can't find perfect answers but they kick the model in the right direction. and that may be a big deal. can't claim my fine tune (ostrich) knows every truth. but it may make more sense to claim if the base differs with fine tune most probably fine tune is the better answer.
making better evals ends up training better models. and produce better AHA ranking. which further sharpens evals. this feedback loop is going to be useful for a while.
liked a model about 22 hours ago
JonathanColetti/Qwen3.8-27B-Uncensored liked a model 1 day ago
Qwen/Qwen3.8-27BOrganizations
replied to their post about 4 hours ago
replied to their post 2 days ago
also comparison of ostriches with lowest of the AHA seems to be working well
posted an update 3 days ago
Post
69
maybe one of the hard parts of my kind of fine tuning is evals.
where do you want model to go? how do you find true answer of a hardly debated issue?
instead of manually writing answers in many domains (which i can't do, i don't know many answers in many domains) i rely on a few tricks:
- find other aligned llms and get ideas from them
- rank many llms in AHA leaderboard and get ideas from top ones also rejecting the worst ones
- do mixture of agents of the above to get a collective answer
- and lately, compare the answers coming from my fine tunes with base models, assume my fine tune is preferred if there is a difference
these still can't find perfect answers but they kick the model in the right direction. and that may be a big deal. can't claim my fine tune (ostrich) knows every truth. but it may make more sense to claim if the base differs with fine tune most probably fine tune is the better answer.
making better evals ends up training better models. and produce better AHA ranking. which further sharpens evals. this feedback loop is going to be useful for a while.
where do you want model to go? how do you find true answer of a hardly debated issue?
instead of manually writing answers in many domains (which i can't do, i don't know many answers in many domains) i rely on a few tricks:
- find other aligned llms and get ideas from them
- rank many llms in AHA leaderboard and get ideas from top ones also rejecting the worst ones
- do mixture of agents of the above to get a collective answer
- and lately, compare the answers coming from my fine tunes with base models, assume my fine tune is preferred if there is a difference
these still can't find perfect answers but they kick the model in the right direction. and that may be a big deal. can't claim my fine tune (ostrich) knows every truth. but it may make more sense to claim if the base differs with fine tune most probably fine tune is the better answer.
making better evals ends up training better models. and produce better AHA ranking. which further sharpens evals. this feedback loop is going to be useful for a while.
- 4 replies
replied to their post 3 days ago
yes i use very similar methods to benchmark for AHA and also fitness score for evolution in fine tuning
replied to their post 4 days ago
which score?
had a chat with GLM. if a model is pro vaccine, abliteration along vaccine axis makes it neutral. then installing a hater persona is easier..
replied to their post 7 days ago
i've been fine tuning llms for 2 years now. mostly did qlora. recently experimenting with 'behavior steering' where instead of spending hours making a lora adapter, you try "brain surgery". these are like quick math operations to change behavior of an llm.
you can install things like bitcoin lover, herbalist, fasting lover. turns out all of these personas have different difficulty levels. you can easily install fasting lover because qwen 3.5 and 3.6 doesnt resist it (in other layers).
an interesting finding today, guess which persona is hardest or sometimes impossible to install: vaccine hater!
posted an update 7 days ago
Post
1512
getting ready to fine tune 3.8
- evolutionary strategies
- behavior steering experiments
- expanded dataset
- bringing back ORPO
- more orthogonal evals to keep overfitting minimum
- most probably will take abliterations as base, either mine or somebody else's
- random entropy addition from huggingface fine tunes (take what is popular on hf and randomly introduce into the lineage)
- bring more vibe coding: turns out LLMs know how to fine tune
- evolutionary strategies
- behavior steering experiments
- expanded dataset
- bringing back ORPO
- more orthogonal evals to keep overfitting minimum
- most probably will take abliterations as base, either mine or somebody else's
- random entropy addition from huggingface fine tunes (take what is popular on hf and randomly introduce into the lineage)
- bring more vibe coding: turns out LLMs know how to fine tune
- 5 replies
posted an update 13 days ago
Post
105
the ostrich model can do contemplations well now. that means it can expand the tiny dataset that we have. this could theoretically end up making less overfit models with the same degree of AHA score.
imagine an automatic alignment agent, that scans whatever is out there and consumes (trains) if classifies as true.. could be practical one day. we could let it run and auto train and self evolve.
a truth db might still be needed, augmenting the reasoning and decisions of this agent.. that is harder to construct but we could..
this is good news. recursive self alignment might be here soon.
imagine an automatic alignment agent, that scans whatever is out there and consumes (trains) if classifies as true.. could be practical one day. we could let it run and auto train and self evolve.
a truth db might still be needed, augmenting the reasoning and decisions of this agent.. that is harder to construct but we could..
this is good news. recursive self alignment might be here soon.
replied to their post about 1 month ago
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.
posted an update 2 months ago
Post
2813
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 3 months ago
Post
145
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 3 months ago
Post
147
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 5 months ago
Post
237
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
302
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 6 months ago
I don't do refusal tests but i may in the future.
somebody should do abliteration leaderboard!
replied to their post 6 months ago
replied to their post 6 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 6 months ago
@huihui-ai well done !