README.md

---
license: apache-2.0
---

This repository contains 2-bit quantized LLaMA-v1 models in GGUF format for use with [llama.cpp](https://github.com/ggerganov/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.cpp` `Q2_K` quantization 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 |