File size: 690 Bytes
3d1dca3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
---
library_name: transformers
language:
- en
pipeline_tag: text-generation
tags:
- nlp
- code
- microsoft
---
Official [AQLM](https://arxiv.org/abs/2401.06118) quantization of [microsoft/Phi-3-mini-128k-instruct
](https://huggingface.co/microsoft/Phi-3-mini-128k-instruct).

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

Results: 
| Model      | Quantization | MMLU (5-shot) | ArcC| ArcE| Hellaswag | Winogrande | PiQA | Model size, Gb |
|------|------|-------|------|------|------|------|------|------|
| microsoft/Phi-3-mini-128k-instruct| None | 0.6881	| 0.5418 | 0.8127 | 0.5980	| 0.7873 | 0.7340 | 7.6 |
|  | 1x16 | 0.5815	| 0.4599 | 0.7845 | 0.5235	| 0.7666 | 0.6930 | 1.4 |