Update README.md
Browse files
README.md
CHANGED
@@ -12,6 +12,9 @@ license_link: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/blob/m
|
|
12 |
ν₯ν μ°κ΅¬μμλ μ΄λ¬ν μ§μμ΄λ κΈ°μ μ΄ μ¬λμ μ νΈμ μ‘°νλ μ μλλ‘ [Direct Preference Optimization](https://arxiv.org/abs/2305.18290)κ³Ό κ°μ κΈ°λ²μ ν΅ν΄ νμ΅ν κ³νμ
λλ€.<br><br>
|
13 |
νμ¬ μ
λ‘λλ μμ€ μ½λμ κ²½μ° μΈκ° μ νΈ λ°μ΄ν°μ
λ§ κ΅¬μΆ λλ€λ©΄ μ κΈ°λ²μ μ μ©ν μ μλλ‘ μμ±λμμΌλ©°, μΆν μλμΌλ‘ μ μ©λμ΄ νμ΅λ μ μλλ‘ κ³ λν μμ μ
λλ€.
|
14 |
|
|
|
|
|
|
|
15 |
## Quick Start
|
16 |
```python
|
17 |
import torch
|
@@ -57,41 +60,6 @@ summary = tokenizer.decode(outputs[0][source.shape[-1]:], skip_special_tokens=Tr
|
|
57 |
"""
|
58 |
```
|
59 |
|
60 |
-
## νμ΅ λ° νκ° λ°©λ²
|
61 |
-
|
62 |
-
### μ€λΉ
|
63 |
-
|
64 |
-
```
|
65 |
-
# νμ λΌμ΄λΈλ¬λ¦¬λ₯Ό μ€μΉν©λλ€.
|
66 |
-
pip install -r requirements.txt
|
67 |
-
```
|
68 |
-
|
69 |
-
```
|
70 |
-
# νμ΅ λ° νκ° λ°μ΄ν°μ
μ μμΉμν΅λλ€.
|
71 |
-
dcs_2024_data
|
72 |
-
βββ μΌμλνμμ½_train.json
|
73 |
-
βββ μΌμλνμμ½_dev.json
|
74 |
-
βββ μΌμλνμμ½_train.json
|
75 |
-
```
|
76 |
-
|
77 |
-
huggingfaceμμ [EXAONE](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct) access tokenμ [λ°κΈ](https://huggingface.co/docs/hub/security-tokens)λ°μ΅λλ€.
|
78 |
-
|
79 |
-
### νμ΅
|
80 |
-
|
81 |
-
```
|
82 |
-
CUDA_VISIBLE_DEVICES=0,1,2 python -m train model=exaone datasets=[DCS] loss=sft exp_name=exaone_sft batch_size=4 max_prompt_length=2048 max_length=2560 token='{access_token}'
|
83 |
-
```
|
84 |
-
|
85 |
-
νμ΅μ΄ μλ£λλ©΄ `model_ckpt/root/` λλ ν 리μμ `exp_name + {νμ΅ μμ μκ°}`μ ν΄λΉνλ λͺ¨λΈ 체ν¬ν¬μΈνΈ νμΌμ΄ μ μ₯λ©λλ€.
|
86 |
-
|
87 |
-
### μΆλ‘
|
88 |
-
|
89 |
-
```
|
90 |
-
python inference/run_test.py --output result.json --model_id model_ckpt/root/exaone_sft_2024-08-26_16-28-10_430889/step-992/ --device cuda:0
|
91 |
-
```
|
92 |
-
|
93 |
-
`model_id` λ³μμ λͺ¨λΈ 체ν¬ν¬μΈνΈ κ²½λ‘λ₯Ό μ
λ ₯ν λ€ μ€ννλ©΄ μμ€μ½λ ν΄λμ `output` νμΌμ΄ μμ±λκ³ , ν΄λΉ νμΌ μ μΆ μ μμν(리λ보λ)μ μ±μ μ΄ λ°μλ©λλ€.
|
94 |
-
|
95 |
## Citing
|
96 |
```
|
97 |
@inproceedings{exaone_sft_dcs,
|
|
|
12 |
ν₯ν μ°κ΅¬μμλ μ΄λ¬ν μ§μμ΄λ κΈ°μ μ΄ μ¬λμ μ νΈμ μ‘°νλ μ μλλ‘ [Direct Preference Optimization](https://arxiv.org/abs/2305.18290)κ³Ό κ°μ κΈ°λ²μ ν΅ν΄ νμ΅ν κ³νμ
λλ€.<br><br>
|
13 |
νμ¬ μ
λ‘λλ μμ€ μ½λμ κ²½μ° μΈκ° μ νΈ λ°μ΄ν°μ
λ§ κ΅¬μΆ λλ€λ©΄ μ κΈ°λ²μ μ μ©ν μ μλλ‘ μμ±λμμΌλ©°, μΆν μλμΌλ‘ μ μ©λμ΄ νμ΅λ μ μλλ‘ κ³ λν μμ μ
λλ€.
|
14 |
|
15 |
+
## νμ΅ λ° νκ° λ°©λ²
|
16 |
+
https://github.com/BM-K/2024-NIKL-DCS
|
17 |
+
|
18 |
## Quick Start
|
19 |
```python
|
20 |
import torch
|
|
|
60 |
"""
|
61 |
```
|
62 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
## Citing
|
64 |
```
|
65 |
@inproceedings{exaone_sft_dcs,
|