Commit
β’
24d656d
1
Parent(s):
4164e46
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,84 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- ko
|
4 |
+
library_name: transformers
|
5 |
+
pipeline_tag: text-generation
|
6 |
+
license: cc-by-nc-sa-4.0
|
7 |
+
---
|
8 |
+
**The license is `cc-by-nc-sa-4.0`.**
|
9 |
+
|
10 |
+
# **π»ββοΈCOKAL_merged_test-v1-13Bπ»ββοΈ**
|
11 |
+
![img](./COKAL-DPO_bear.png)
|
12 |
+
|
13 |
+
## Model Details
|
14 |
+
|
15 |
+
**Model Developers** Seungyoo Lee(DopeorNope)
|
16 |
+
|
17 |
+
**Input** Models input text only.
|
18 |
+
|
19 |
+
**Output** Models generate text only.
|
20 |
+
|
21 |
+
**Model Architecture**
|
22 |
+
COKAL_merged_test-v1-13B is an auto-regressive language model based on the LLaMA2 transformer architecture.
|
23 |
+
|
24 |
+
|
25 |
+
**Base Model**
|
26 |
+
|
27 |
+
[HumanF-MarkrAI/COKAL-DPO-13b-v2](https://huggingface.co/HumanF-MarkrAI/COKAL-DPO-13b-v2)
|
28 |
+
|
29 |
+
[MarkrAI/DopeorNope-maestro-v2-DPO-13b](https://huggingface.co/MarkrAI/DopeorNope-maestro-v2-DPO-13b)
|
30 |
+
|
31 |
+
# **Implemented Method**
|
32 |
+
|
33 |
+
I utilized `slerp merging` to smoothly blend the gradients of the base models to create my model.
|
34 |
+
|
35 |
+
The merging approach relies on luck, but at the same time, if I have an accurate understanding of my model's performance, I can carefully select models that excel in each aspect to develop a well-balanced model.
|
36 |
+
|
37 |
+
|
38 |
+
|
39 |
+
# **Model Benchmark**
|
40 |
+
|
41 |
+
|
42 |
+
## KO-LLM leaderboard
|
43 |
+
- Follow up as [Open KO-LLM LeaderBoard](https://huggingface.co/spaces/upstage/open-ko-llm-leaderboard).
|
44 |
+
|
45 |
+
| Model | Average |Ko-ARC | Ko-HellaSwag | Ko-MMLU | Ko-TruthfulQA | Ko-CommonGen V2 |
|
46 |
+
| --- | --- | --- | --- | --- | --- | --- |
|
47 |
+
| COKAL_merged_test-v1-13Bπ»ββοΈ | 52.72 | 51.45 | 60.55 | 44.8 | 49.05 | 57.73 |
|
48 |
+
| [COKAL-DPO-13b-v2π»ββοΈ](https://huggingface.co/HumanF-MarkrAI/COKAL-DPO-13b-v2) | 52.69 | 54.95 | 63.02 | 43.98 | 51.67 | 49.82 |
|
49 |
+
| [COKAL-DPO_test-v2-13bπ»ββοΈ](https://huggingface.co/DopeorNope/COKAL-DPO_test-v2-13b) | 52.67 | 55.63 | 63.5 | 43.49 | 51.5 | 49.23 |
|
50 |
+
| hyeogi/Yi-6b-dpo-v0.2](https://huggingface.co/hyeogi/Yi-6b-dpo-v0.2) | 52.63 | 41.72 | 52.96 | 46.69 | 52.38 | 69.42 |
|
51 |
+
| [DopeorNope-maestro-v2-DPO-13bπ»ββοΈ](https://huggingface.co/MarkrAI/DopeorNope-maestro-v2-DPO-13b) | 49.42 | 45.14 | 56.69 | 41.37 | 42.26 | 61.63 |
|
52 |
+
|
53 |
+
|
54 |
+
|
55 |
+
# Implementation Code
|
56 |
+
|
57 |
+
## Load model
|
58 |
+
```python
|
59 |
+
|
60 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
61 |
+
import torch
|
62 |
+
|
63 |
+
repo = "DopeorNope/COKAL_merged_test-v1-13B"
|
64 |
+
OpenOrca = AutoModelForCausalLM.from_pretrained(
|
65 |
+
repo,
|
66 |
+
return_dict=True,
|
67 |
+
torch_dtype=torch.float16,
|
68 |
+
device_map='auto'
|
69 |
+
)
|
70 |
+
OpenOrca_tokenizer = AutoTokenizer.from_pretrained(repo)
|
71 |
+
```
|
72 |
+
## Prompt (Alpaca format)
|
73 |
+
|
74 |
+
```python
|
75 |
+
|
76 |
+
prompt= f"μλλ λ¬Έμ λ₯Ό μ€λͺ
νλ μ§μμ¬νκ³Ό, ꡬ체μ μΈ λ΅λ³μ λ°©μμ μꡬνλ μ
λ ₯μ΄ ν¨κ» μλ λ¬Έμ₯μ
λλ€. μ΄ μμ²μ λν΄ μ μ νκ² λ΅λ³ν΄μ£ΌμΈμ.\n\n### μ§μμ¬ν:\n{instruction}\n\n### μ
λ ₯:\n{input}\n\n### λ΅λ³:\n"
|
77 |
+
|
78 |
+
prompt_no_input = f"μλλ λ¬Έμ λ₯Ό μ€λͺ
νλ μ§μμ¬νμ
λλ€. μ΄ μμ²μ λν΄ μ μ νκ² λ΅λ³ν΄μ£ΌμΈμ.\n\n### μ§μμ¬ν:\n{instruction}\n\n### λ΅λ³:\n"
|
79 |
+
|
80 |
+
|
81 |
+
```
|
82 |
+
|
83 |
+
|
84 |
+
---
|