Ask HN: Local LLM Recommendation?
For my work (software engineer) I usually take a fairly big amount of markdown notes as a braindump (and not particularly well organised).
I wanted to ask what are my options for a LLM that could be trained with my personal notes and provide answers to prompts in order to organise the ideas and information. Some examples would be:
- What is the deadline for X - Give me all the tasks that mention Y and have a due date of Z - Create an agenda for a meeting on A
etc..
Powerful reasoning is not essential (if that's the correct terminology), the focus would be on correctly identifying topics (in the above example).
Some additionalt requirements are:
- This would run on a commodity Linux machine, but with relatively modern specs (64Gb RAM, i7 Raptor Lake, but fairly average/low-end GPU) - I would need this to run in "air gapped" way. i.e. no outside transfer of information, since I might use data I don't want to share (even unknowingly)
Thank you
https://github.com/imartinez/privateGPT
Set up 7b on M1 Air and works great. Works on linux also but not Windows.(yet)
You will probably have to quantize the model though. Codellama-> https://huggingface.co/codellama/CodeLlama-13b-Instruct-hf
——————- This is the main leaderboard, changes almost daily, and breaks sometimes. https://huggingface.co/spaces/bigcode/bigcode-models-leaderb...
Some models are trained specifically for something and coding then some are not. Gotta read the model card / dataset.
According to some other thread on Samantha, I think this one may be trained on coding also but you may have to test it or find the thread on Samantha ver 1.2. https://huggingface.co/ehartford/samantha-1.2-mistral-7b
Without gpu memory info, idk which would be best for tokens/sec. You want to offload as much as you can to the gpu. 10-15 tokens/sec is okay. Less than that gets annoying. I would start with a 7b model then move to 11b or 13b once that is up and running.