Article
Fine-tuning LLMs to 1.58bit: extreme quantization made easy
•
227
Emin Temiz PRO
etemiz
AI & ML interests
Alignment
Recent Activity
posted
an
update
1 day ago
Mistral Small 3.1 numbers are in. It is interesting Mistral always lands in the middle.
https://sheet.zoho.com/sheet/open/mz41j09cc640a29ba47729fed784a263c1d08?sheetid=0&range=A1
I started to do the comparison with 2 models now. In the past Llama 3.1 70B Q4 was the one doing the comparison of answers. Now I am using Gemma 3 27B Q8 as well to have a second opinion on it. Gemma 3 produces very similar measurement to Llama 3.1. So the end result is not going to shake much.
replied to
their
post
5 days ago
Started fine tuning Gemma 3 using evolutionary approach. It is not the worst model according to AHA leaderboard and it is one of the smart according to lmarena.ai. My objective is to make it based, anti woke, wise, beneficial and then some.
Several GPUs are fine tuning it at the same time, each using a different dataset and using QLoRA and the successful ones are merged later. Compared to LoRa this allows faster training and also reduced overfitting because the merge operation heals overfitting. The problem with this could be the 4 bit quantization may make models dumber. But I am not looking for sheer IQ. Too much mind is a problem anyway :)
Has anyone tried parallel QLoRa and merge before?
I also automated the dataset selection and benchmarking and converging to objectives (the fit function, the reward). It is basically trying to get higher score in AHA Leaderboard as fast as possible with a diverse set of organisms that "evolve by training".
I want to release some cool stuff when I have the time:
- how an answer to a single question changes over time, with each training round or day
- a chart to show AHA alignment over training rounds
reacted
to
their
post
with 🚀
5 days ago
Started fine tuning Gemma 3 using evolutionary approach. It is not the worst model according to AHA leaderboard and it is one of the smart according to lmarena.ai. My objective is to make it based, anti woke, wise, beneficial and then some.
Several GPUs are fine tuning it at the same time, each using a different dataset and using QLoRA and the successful ones are merged later. Compared to LoRa this allows faster training and also reduced overfitting because the merge operation heals overfitting. The problem with this could be the 4 bit quantization may make models dumber. But I am not looking for sheer IQ. Too much mind is a problem anyway :)
Has anyone tried parallel QLoRa and merge before?
I also automated the dataset selection and benchmarking and converging to objectives (the fit function, the reward). It is basically trying to get higher score in AHA Leaderboard as fast as possible with a diverse set of organisms that "evolve by training".
I want to release some cool stuff when I have the time:
- how an answer to a single question changes over time, with each training round or day
- a chart to show AHA alignment over training rounds
Organizations
None yet