MugoSquero
commited on
Commit
•
208599c
1
Parent(s):
39028a3
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
pipeline_tag: text-generation
|
3 |
+
tags:
|
4 |
+
- phi-msft
|
5 |
+
language:
|
6 |
+
- en
|
7 |
+
library_name: transformers
|
8 |
+
---
|
9 |
+
# LM-Cocktail phi-2 v1.1
|
10 |
+
|
11 |
+
This is a 0.5-0.5 merge of two models based on phi-2. Here are the models used to create this merge:
|
12 |
+
1. [venkycs/phi-2-instruct](https://huggingface.co/venkycs/phi-2-instruct)
|
13 |
+
2. [Yhyu13/phi-2-sft-dpo-gpt4_en-ep1](https://huggingface.co/Yhyu13/phi-2-sft-dpo-gpt4_en-ep1)
|
14 |
+
|
15 |
+
I named this model "LMCocktail phi-2 v1.1" because I see it as a continuation of the [v1](https://huggingface.co/Yhyu13/LMCocktail-phi-2-v1).
|
16 |
+
|
17 |
+
I used [Yhyu13/phi-2-sft-dpo-gpt4_en-ep1](https://huggingface.co/Yhyu13/phi-2-sft-dpo-gpt4_en-ep1) and it "outputs significantly longer result" than the one used in v1 by Yhyu13.
|
18 |
+
|
19 |
+
I also used [venkycs/phi-2-instruct](https://huggingface.co/venkycs/phi-2-instruct) "a fine-tuned version of [microsoft/phi-2](https://huggingface.co/microsoft/phi-2) on the filtered [ultrachat200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) dataset using the SFT technique".
|
20 |
+
|
21 |
+
The main reason I created this model was to merge it with [cognitivecomputations/dolphin-2_6-phi-2](https://huggingface.co/cognitivecomputations/dolphin-2_6-phi-2), and I will create a repo for it when I do it.
|
22 |
+
|
23 |
+
# Code
|
24 |
+
|
25 |
+
The LM-cocktail is novel technique for merging multiple models: https://arxiv.org/abs/2311.13534
|
26 |
+
|
27 |
+
Code is backed up by this repo: https://github.com/FlagOpen/FlagEmbedding.git
|
28 |
+
|
29 |
+
Merging script is available under the [./scripts](./scripts) folder.
|