Thanks for gguf version of my model (SimonGPT-simple-instruct)
Thanks for making a GGUF version of my model, i wanna say it would be great if you could inclode in the readme that the model works best with temp around 0.01, also i would be happy if you could say that the model is made for this format:
Prompter: Whats a cat? Assistant:
The ai will auto say the stuff after the assistant: also it works without it but its recomended, if you find any problems in the model i would be happy to try to fix them (: also happy new year!
The model card is unfortunately automatically generated and so any changes to it would get automatically reverted but our convenient download page under includes the official model card https://hf.tst.eu/model#SimonGPT-simple-instruct-GGUF contains the original model card. In addition, we are linking the original model in the model card. I think by mentioning it inside https://huggingface.co/mradermacher/SimonGPT-simple-instruct-GGUF/discussions/1 you already did the best you could to inform the users of your model.
With the model already having reached 245 downloads we probably should generate the better weighted/imatrix quants for it as well. I will now do so.
I am thinking about maybe making some new training data and training it again soon, but i will have to think before that beacose on my pc to train it once it takes ~22h using pytorch
(i have a intel xeon e5-2650 v4 and 32 gigs of ram)
Thanks, also these new imatrix are more efficient and optimized if i know corectly better for future experiments i wanna make :)
I am thinking about maybe making some new training data and training it again soon, but i will have to think before that beacose on my pc to train it once it takes ~22h using pytorch
(i have a intel xeon e5-2650 v4 and 32 gigs of ram)
I could let you use some of my resources for training. I have AMD Ryzen Threadripper PRO 7975WX and 512 GB of DDR5 RDIMM ECC RAM. Can't you just train such a tiny model on the GPU? It should easely fit and training on GPU far faster than training on CPU.
Thanks, also these new imatrix are more efficient and optimized if i know corectly better for future experiments i wanna make :)
No problem. Glad I was able to help you and your users test this model by providing weighted/imatrix quants.
I am thinking about maybe making some new training data and training it again soon, but i will have to think before that beacose on my pc to train it once it takes ~22h using pytorch
(i have a intel xeon e5-2650 v4 and 32 gigs of ram)I could let you use some of my resources for training. I have AMD Ryzen Threadripper PRO 7975WX and 512 GB of DDR5 RDIMM ECC RAM. Can't you just train such a tiny model on the GPU? It should easely fit and training on GPU far faster than training on CPU.
The problem is that my gpu (rx vega 56 8gb) doesnt have cuda if i know corectly, and right now i am using a random python app with torch that ai helped me develop
The problem is that my gpu (rx vega 56 8gb) doesnt have cuda if i know corectly, and right now i am using a random python app with torch that ai helped me develop
I have many GPUs with CUDA support I could let you use. Threadripper has an RTX 2070 super, CastlePeak has an RTX 3080 and StormPeak 2x RTX 4090. I could create you an LXC container on Threadripper where you could run your training. It will likely be around 20 times faster to run your training on GPU and given how tiny your model is doing so will take almost no resources. If you are interested you could add me on Discord with username "nicobosshard". You could probably also just run your training on Google collab as you need such little resources that you fall well withing the limits of their free plan.