File size: 2,029 Bytes
2936c81
 
 
1108a1c
a4563de
1108a1c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: apache-2.0
---

# LM-cocktail Mistral 7B v1


This is a 50%-50% model of two best Mistral models

https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2

https://huggingface.co/xDAN-AI/xDAN-L1-Chat-RL-v1

both claimed to be better than chatgpt-3.5-turbo in almost all metrics.

# Alpaca Eval

I am thrilled to announce that ChatGPT has ranked LMCocktail 7B as the second best model next to GPT4 on AlpcaEval in my local community run, even greater than my previously best [LMCocktail-10.7B-v1](https://huggingface.co/Yhyu13/LMCocktail-10.7B-v1.git) model. You can also check the leaderboard at [./Alpaca_eval/chatgpt_fn_--LMCocktail-Mistral-7B-v1/](./Alpaca_eval/chatgpt_fn_--LMCocktail-Mistral-7B-v1/)

```
                        win_rate  standard_error  n_total  avg_length
gpt4                       73.79            1.54      805        1365
LMCocktail-7B-v1(new)      73.54            1.55      805        1870
LMCocktail-10.7B-v1(new)   73.45            1.56      804        1203
claude                     70.37            1.60      805        1082
chatgpt                    66.09            1.66      805         811
wizardlm-13b               65.16            1.67      805         985
vicuna-13b                 64.10            1.69      805        1037
guanaco-65b                62.36            1.71      805        1249
oasst-rlhf-llama-33b       62.05            1.71      805        1079
alpaca-farm-ppo-human      60.25            1.72      805         803
falcon-40b-instruct        56.52            1.74      805         662
text_davinci_003           50.00            0.00      805         307
alpaca-7b                  45.22            1.74      805         396
text_davinci_001           28.07            1.56      805         296
```


# Code

The LM-cocktail is novel technique for merging multiple models https://arxiv.org/abs/2311.13534

Code is backed up by this repo https://github.com/FlagOpen/FlagEmbedding.git

Merging scripts available under the [./scripts](./scripts) folder