Qwen 3.5 397b and GLM 5.1 Opus fine tune
Posted by No_Farmer_495@reddit | LocalLLaMA | View on Reddit | 2 comments
Hi all. Many models on hugging face have been fine tuned with that 3000x opus dataset, but the two I mentioned in the title are missing it. Could anyone with available compute fine tune them? Or does a similar fine tune of these models already exist??
Charming_Support726@reddit
At first it isn't that cheap to run a training on a model of this size. (try for yourself) An mostly there won't be any ROI.
At second the outcome is, lets call it, questionable. A SFT with this dataset might change the behavior a bit, but won't alter the way it reasons in depth.
Opus gets its distinct behavior by programs of RL style training. At least that is what many people guess. Running a finetune with the traces is like eating a paper with Einstein's theory on it.
No_Farmer_495@reddit (OP)
It was mostly for the style that I wanted the fine tune. It's a shame big models never get fine tunes :(