File size: 2,642 Bytes
e46b821
 
 
507385f
d8d65a6
359cf00
 
c4b41b5
359cf00
c4b41b5
4c9cf74
 
359cf00
f19cae9
359cf00
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f19cae9
58308ba
e0f47d8
 
 
467df4f
 
 
7659f35
16d4d23
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
---
license: apache-2.0
---
<img src="https://huggingface.co/Vezora/Mistral-22B-v0.1/resolve/main/unsloth.png" width="100" height="150" />

## Mistral-22b-V.01 Release Announcement 🚀

## This model is not an moe, it is infact a 22B parameter dense model!

**Date**: April 11
**Creator** Nicolas Mejia-Petit

### Overview
Just one day after the release of **Mixtral-8x-22b**, we are excited to introduce our handcrafted experimental model, **Mistral-22b-V.01**. This model is a culmination of equal knowledge distilled from all experts into a single, dense 22b model. This model is not a single trained expert, rather its a compressed MOE model, turning it into a dense 22b mode. This is the first working MOE to Dense model conversion.

### Capabilities
- **Math Proficiency**: The model exhibits strong mathematical abilities.
- **Reasoning**: It possesses decent reasoning skills.
- **User Interaction**: It responds effectively to user prompts.

### Experimental Nature
Please note that Mistral-22b-V.01 is an experimental model. It has been fine-tuned with fewer examples compared to the model set for release tomorrow. We encourage you to explore its capabilities and provide feedback.

### Upcoming Release: V.2
Stay tuned for the release of **V.2** tomorrow, which will feature enhancements in:
- Multi-turn conversations
- Multiturn coding
- JSON mode
- Agent abilities

### Background
The decision to release this experimental version was prompted by someone attempting to replicate my experiment based on my tweets. We wanted to ensure our community has access to the official version first.

### Stay Updated
Keep an eye out for **V.2**, it's going to be a game-changer! And is currently training, will be done in the next ~24 hours. 🌟Paper Coming Soon🌟

## Thank you!
- Thank you to Daniel Han, for Unsloth AI which was used to train this model. this led to a 2-3x speed increae and 2-3x decrease in memmory consumption.
- Thank you to Charles Coddard, for providng me with a script that was nessary to make this model.
- Thank you to Mistral, for releasing Another Wonderful open source model, under Apache 2.0.
- Thank you to Tim Dettmers, for creating QLora
- Thank you to Tri Dao, for creating Flash Attention
- Thank you to Microsoft, for the Lora paper, and the Slice-GPT paper.
- Thank you to the Hugging Face team, for everything.❤️ We really do appreciate you guys and all your hard work and commitment to the open source community!❤️

## I will answer more questions about this model tomorrow, I have not slept since mixtral 22bx8 dropped. Base model + V.2 Lora checkpoint will be available tomorrow ##