File size: 825 Bytes
4c3d1a4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
90f515b
 
 
ffdffd6
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
---
library_name: transformers
tags:
- llama
- facebook
- meta
- llama-3
- conversational
- text-generation-inference
---
Official [AQLM](https://arxiv.org/abs/2401.06118) quantization of [meta-llama/Meta-Llama-3-8B-Instruct
](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct).

For this quantization, we used 1 codebook of 16 bits.

Results: 
| Model      | Quantization | MMLU (5-shot) | GSM8k (8-shot) | ArcC| ArcE| Hellaswag | Winogrande | PiQA | Model size, Gb |
|------|------|-------|------|------|------|------|------|------|------|
|meta-llama/Meta-Llama-3-8B-Instruct| None |0.6560 | 0.7475 | 0.5299 | 0.8165 | 0.5771 | 0.7867	| 0.7206 | 16.1 |
|  | 1x16 | 0.5872 | 0.5087 | 0.4590 | 0.7710 | 0.5491 | 0.7726 | 0.6953 | 4.1 |

**UPD** 02.05.2024

The version of model with improved fine-tuning procedure.