Are these models trained from scratch, or just different quantizations of the same model?
Firstly, I would like to thank TheBloke for publishing these, it's awesome.
My question is, what is the "value add" here? Are these models any different than Meta's Llama 2?
I am wondering if TheBloke:
- Trained each of these models separately from scratch?
- Or just provides different quantizations of the same pre-trained model?
I guess in general, besides posting these models on huggingface, what other things is TheBloke doing for these models?
The latter. I provide quantisations in various formats and quantisation parameter permutations, along with documentation on how to get them working for various use cases. Then I keep them updated as quantisation technology changes and improves.
Thank you @TheBloke ! One follow up question, where did you get the original weights? Was it directly from a Llama 2 trained by Meta?
The latter. I provide quantisations in various formats and quantisation parameter permutations, along with documentation on how to get them working for various use cases. Then I keep them updated as quantisation technology changes and improves.
The actual MVP.
@TheBloke is not the hero we deserve, but he's the hero we need.
🦇🦇🦇