It has been brought to my attention that perhaps the original non gguf files I uploaded will need to be updated. I apologize for the inconvenience to the community. After all the gguf files are done uploading I will look at the original files and the tokenizer and make sure you can use those to make your own quants. I apologize to all of you for this. I will do better. I am VERY inexperienced with this and, I am trying hard to give you all the files. I will do my best to make this better for the community. Thank you @mradermacher for helping me understand that I did a pretty poor job with those files. I will get better at this.
I changed the optimizer and many of the training settings I am now getting a MUCH lower loss rate. I will continue training to bring the current 0.6 down to a better 0.2 range and I will upload all the updated files. This way we all have a better and MUCH smarter model to work with. The new settings make llama 3.1 8b work amazing with this dataset. I will do my best to get the new version up as fast as possible. The older model with the 0.6 loss is up but the new model will blow this one out the water. :)
Uploading a new set of files with the corrected config files in place today. Some how last upload wiped the config.json. I may just start uploading with the web interface instead of python code.
What I have found is that if you use the merged model then it works like trash. HOWEVER, when I load the lora model instead... I get wonderful results and its great. I will work on uploading the lora I created shortly.
I am pretty new to uploading models. I think I made an error when I loaded my model from unsloth I loaded a 4bit then I saved to 16bit this is why the lora works and not the model its self. I will retrain this and upload new files asap.
