This repository contains 2-bit quantized LLaMA-v1 models in GGUF format for use with llama.cpp.
All tensors are quantized with Q2_K, except for output.weight, which is Q6_K, and, in the case of LLaMA-v2-70B, attn_v, which is Q4_K.
The quantized models differ from the standard llama.cpp 2-bit quantization in two ways:
- These are actual 2-bit quantized models instead of the mostly 3-bit quantization provided by the standard
llama.cppQ2_Kquantization method - The models were prepared with a refined (but not yet published) k-quants quantization approach
The table shows Wikitext perplexities for a context length of 2048 tokens computed with these models using llama.cpp
| Model | Perplexity |
|---|---|
| 7B | 6.4023 |
| 13B | 5.3967 |
| 30B | 4.5065 |
| 65B | 3.9136 |
- Downloads last month
- 15