Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
model-index:
|
3 |
+
- name: EEVE-Math-10.8B-SFT
|
4 |
+
results:
|
5 |
+
- task:
|
6 |
+
type: text-generation
|
7 |
+
dataset:
|
8 |
+
name: gsm8k-ko
|
9 |
+
type: gsm8k
|
10 |
+
metrics:
|
11 |
+
- name: pass@1
|
12 |
+
type: pass@1
|
13 |
+
value: 0.539
|
14 |
+
verified: false
|
15 |
+
base_model: yanolja/EEVE-Korean-10.8B-v1.0
|
16 |
+
---
|
17 |
+
# EEVE-Math-10.8B-SFT
|
18 |
+
์ด ๋ชจ๋ธ์ [Orca-Math: Unlocking the potential of SLMs in Grade School Math](https://arxiv.org/pdf/2402.14830.pdf)๋ฐ [DARE](https://arxiv.org/abs/2311.03099)์ ๊ฐ๋
๊ณผ ์ด๋ฅผ ํ์ฉํ ๋ด์ฉ์ ํฌํจํ๊ณ ์์ต๋๋ค.
|
19 |
+
|
20 |
+
| Model | gsm8k(pass@1) |
|
21 |
+
|---|---|
|
22 |
+
| Base | 0.4049 |
|
23 |
+
| SFT Epoch 1 | 0.508 |
|
24 |
+
| SFT Epoch 2(M1) | **0.539** |
|
25 |
+
| SFT -> KTO(M2) | - |
|
26 |
+
| SFT -> KTO -> KTO(Final) | - |
|
27 |
+
|
28 |
+
|
29 |
+
## Specifications
|
30 |
+
SFT(M1) ๋จ๊ณ
|
31 |
+
|
32 |
+
## Base Model
|
33 |
+
[yanolja/EEVE-Korean-10.8B-v1.0](https://huggingface.co/yanolja/EEVE-Korean-10.8B-v1.0)
|
34 |
+
|
35 |
+
## Dataset
|
36 |
+
[orca-math-word-problems-193k-korean](https://huggingface.co/datasets/kuotient/orca-math-word-problems-193k-korean)
|
37 |
+
|
38 |
+
## Evaluation
|
39 |
+
[gsm8k-ko](https://huggingface.co/datasets/kuotient/gsm8k-ko), kobest
|
40 |
+
```
|
41 |
+
git clone https://github.com/kuotient/lm-evaluation-harness.git
|
42 |
+
cd lm-evaluation-harness
|
43 |
+
pip install -e .
|
44 |
+
```
|
45 |
+
```
|
46 |
+
lm_eval --model hf \
|
47 |
+
--model_args pretrained=yanolja/EEVE-Korean-Instruct-2.8B-v1.0 \
|
48 |
+
--tasks gsm8k-ko \
|
49 |
+
--device cuda:0 \
|
50 |
+
--batch_size auto:4
|
51 |
+
```
|
52 |
+
|
53 |
+
| Model | gsm8k(pass@1) | boolq(acc) | copa(acc) | hellaswag(acc) | Overall |
|
54 |
+
|---|---|---|---|---|---|
|
55 |
+
| yanolja/EEVE-Korean-10.8B-v1.0 | 0.4049 | - | - | - | - | - |
|
56 |
+
| yanolja/EEVE-Korean-Instruct-10.8B-v1.0 | 0.4511 | **0.8668** | **0.7450** | **0.4940** | 0.6392 |
|
57 |
+
| **EEVE-Math-10.8B-SFT** | **0.5390** | 0.8027 | 0.7260 | 0.4760 | 0.6359 |
|
58 |
+
| **EEVE-Instruct-Math-10.8B** | 0.4951 | 0.8283 | 0.7500 | 0.4880 | **0.640** |
|