|
--- |
|
license: apache-2.0 |
|
--- |
|
|
|
This repository contains alternative Mistral-instruct-7B (https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2) quantized models in GGUF format for use with `llama.cpp`. |
|
The models are fully compatible with the oficial `llama.cpp` release and can be used out-of-the-box. |
|
|
|
I'm carefull to say "alternative" rather than "better" or "improved" as I have not put any effort into evaluating performance |
|
differences in actual usage. Perplexity is lower compared to the "official" `llama.cpp` quantization, but perplexity is not |
|
necessarily a good measure for real world performance. Nevertheless, perplexity does measure quantization error, so below is a table |
|
comparing perplexities of these quantized models to the current `llama.cpp` quantization approach on Wikitext for a context length of 512 tokens. |
|
The "Quantization Error" columns in the table are defined as `(PPL(quantized model) - PPL(fp16))/PPL(fp16)`. |
|
|
|
| Quantization | Model file | PPL(llama.cpp) | Quantization Error | PPL(new quants) | Quantization Error | |
|
|--:|--:|--:|--:|--:|--:| |
|
|Q3_K_S| mistral-instruct-7b-q3k-small.gguf | 6.9959 | 4.27% | 6.8920 | 2.72% | |
|
|Q3_K_M| mistral-instruct-7b-q3k-medium.gguf| 6.8892 | 2.68% | 6.8089 | 1.48% | |
|
|Q4_K_S| mistral-instruct-7b-q4k-small.gguf | 6.7649 | 0.82% | 6.7351 | 0.38% | |
|
|Q5_K_S| mistral-instruct-7b-q5k-small.gguf | 6.7197 | 0.15% | 6.7186 | 0.13% | |
|
|Q4_0 | mistral-instruct-7b-q40.gguf | 6.7728 | 0.94% | 6.7191 | 0.14% | |
|
|
