I'm sorry but this one is a mess at coding or is it just me?
Seems to do pretty well at first when coding but it always misses many small silly things, it reminds for sure about an earlier .gguf version of GLM 4.7 flash.
The ggufs were made using llama.cpp's llama-quantize. Personally I use vllm so I wouldn't really know if there are gguf issues with this model.
I know others have had success with lowering temperature closer to 0.5-0.6, hopefully this helps.
for coding not so much. was worth a shot though, thanks!
Makes sense. Not any data in the dataset to advance or reinforce agentic coding capability. Most likely regressed in those areas
Yeah, maybe you could train on the Minimax 8800x coding dataset than merge the LORAs at like 0.7 to 0.3 to regain some coding without wiping out the reasoning
Yea the next one will be tuned with both the minimax 8800x and the agentic code sft with less aggressive settings