Upload README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,73 @@
|
|
1 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
license: cc-by-nc-sa-4.0
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
language:
|
3 |
+
- ko
|
4 |
+
datasets:
|
5 |
+
- kyujinpy/KOR-OpenOrca-Platypus-v3
|
6 |
+
library_name: transformers
|
7 |
+
pipeline_tag: text-generation
|
8 |
license: cc-by-nc-sa-4.0
|
9 |
---
|
10 |
+
|
11 |
+
# **⭐My custom LLM 13B⭐**
|
12 |
+
|
13 |
+
## Model Details
|
14 |
+
**Model Developers**
|
15 |
+
- Kyujin Han (kyujinpy)
|
16 |
+
|
17 |
+
**Model Architecture**
|
18 |
+
- My custom LLM 13B is an auto-regressive language model based on the LLaMA2 transformer architecture.
|
19 |
+
|
20 |
+
**Base Model**
|
21 |
+
- [beomi/llama-2-koen-13b](https://huggingface.co/beomi/llama-2-koen-13b)
|
22 |
+
|
23 |
+
**Training Dataset**
|
24 |
+
- [kyujinpy/KOR-OpenOrca-Platypus-v3](https://huggingface.co/datasets/kyujinpy/KOR-OpenOrca-Platypus-v3).
|
25 |
+
|
26 |
+
---
|
27 |
+
# Model comparisons
|
28 |
+
> Ko-LLM leaderboard(11/27; [link](https://huggingface.co/spaces/upstage/open-ko-llm-leaderboard))
|
29 |
+
|
30 |
+
| Model | Average | Ko-ARC | Ko-HellaSwag | Ko-MMLU | Ko-TruthfulQA | Ko-CommonGen V2 |
|
31 |
+
| --- | --- | --- | --- | --- | --- | --- |
|
32 |
+
| ⭐My custom LLM 13B-v1⭐ | **50.19** | **45.99** | 56.93 | **41.78** | 41.66 | **64.58** |
|
33 |
+
| ⭐My custom LLM 13B-v2⭐ | 48.28 | 45.73 | **56.97** | 38.77 | 38.75 | 61.16 |
|
34 |
+
| ⭐My custom LLM 13B-v3⭐ | 46.40 | 44.71 | 56.89 | 40.86 | **44.22** | 45.34 |
|
35 |
+
| **⭐My custom LLM 13B-v4⭐** | NaN | NaN | NaN | NaN | NaN | NaN |
|
36 |
+
|
37 |
+
---
|
38 |
+
# Model comparisons2
|
39 |
+
> AI-Harness evaluation; [link](https://github.com/Beomi/ko-lm-evaluation-harness)
|
40 |
+
|
41 |
+
| Model | Copa | Copa | HellaSwag | HellaSwag | BoolQ | BoolQ | Sentineg | Sentineg |
|
42 |
+
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
|
43 |
+
| | 0-shot | 5-shot | 0-shot | 5-shot | 0-shot | 5-shot | 0-shot | 5-shot |
|
44 |
+
| ⭐My custom LLM 13B-v1⭐ | 0.7987 | 0.8269 | 0.4994 | 0.5660 | 0.3343 | 0.5060 | **0.6984** | 0.9723 |
|
45 |
+
| ⭐My custom LLM 13B-v2⭐ | 0.7938 | 0.8209 | 0.4978 | 0.4893 | 0.3343 | 0.5614 | 0.6283 | 0.9773 |
|
46 |
+
| ⭐My custom LLM 13B-v3⭐ | **0.8107** | 0.8359 | **0.5176** | 0.5182 | **0.6702** | 0.7851 | 0.5241 | 0.9698 |
|
47 |
+
| **⭐My custom LLM 13B-v4⭐** | NaN | NaN | NaN | NaN | NaN | NaN | NaN | NaN |
|
48 |
+
| [beomi/llama-2-koen-13b](https://huggingface.co/beomi/llama-2-koen-13b) | 0.7768 | 0.8128 | 0.4999 | 0.5127 | 0.3988 | 0.7038 | 0.5870 | 0.9748 |
|
49 |
+
|
50 |
+
---
|
51 |
+
# Implementation Code
|
52 |
+
```python
|
53 |
+
### KO-Platypus
|
54 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
55 |
+
import torch
|
56 |
+
|
57 |
+
repo = "PracticeLLM/Custom-KoLLM-13B-v4"
|
58 |
+
OpenOrca = AutoModelForCausalLM.from_pretrained(
|
59 |
+
repo,
|
60 |
+
return_dict=True,
|
61 |
+
torch_dtype=torch.float16,
|
62 |
+
device_map='auto'
|
63 |
+
)
|
64 |
+
OpenOrca_tokenizer = AutoTokenizer.from_pretrained(repo)
|
65 |
+
```
|
66 |
+
|
67 |
+
---
|
68 |
+
# Hyperparameters
|
69 |
+
- learning_rate: 4e-4
|
70 |
+
- batch_size: 16
|
71 |
+
- epoch: 1
|
72 |
+
- lora_target_modules: [gate_proj, down_proj, up_proj, q_proj, k_proj, v_proj, o_proj]
|
73 |
+
- cutoff_len: 4096
|