aripos1 commited on
Commit
86de414
Β·
verified Β·
1 Parent(s): 51a54f9

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +32 -57
README.md CHANGED
@@ -4,65 +4,40 @@ datasets:
4
  - aripos1/gorani_dataset
5
  language:
6
  - ko
 
 
7
  base_model:
8
  - unsloth/Llama-3.2-3B-Instruct-bnb-4bit
9
  pipeline_tag: text-generation
10
  library_name: transformers
11
  ---
12
- # Gorani LoRA Merged (Llama 3.2-3B 기반)
13
-
14
- ## πŸ“Œ Model Overview
15
- - **Model Name**: Gorani LoRA Merged
16
- - **Base Model**: `unsloth/Llama-3.2-3B-Instruct-bnb-4bit`
17
- - **Adapter Type**: LoRA
18
- - **Task**: Text Generation (λ²ˆμ—­, 챗봇)
19
- - **License**: Apache 2.0
20
- - **Author**: aripos1
21
- - **Fine-Tuned Dataset**: [Custom Korean-English Dataset]
22
-
23
- ## πŸ›  How to Use
24
- ```python
25
- from transformers import AutoModelForCausalLM, AutoTokenizer
26
-
27
- model_name = "aripos1/gorani-lora-merged"
28
- model = AutoModelForCausalLM.from_pretrained(model_name)
29
- tokenizer = AutoTokenizer.from_pretrained(model_name)
30
-
31
- text = "Translate this sentence from Korean to English: μ•ˆλ…•ν•˜μ„Έμš”."
32
- inputs = tokenizer(text, return_tensors="pt")
33
- outputs = model.generate(**inputs)
34
- print(tokenizer.decode(outputs[0], skip_special_tokens=True))
35
- πŸ“Š Model Performance
36
- Version Comet Score BERT Score
37
- v1 0.78 0.85
38
- v2 0.82 0.88
39
- v3 0.85 0.90
40
- πŸ“‚ Training Details
41
- LoRA 적용 μ—¬λΆ€: βœ…
42
- Dataset: Custom dataset (Korean-English parallel corpus)
43
- Frameworks: Hugging Face transformers, peft
44
- Training Hardware: NVIDIA A100 40GB
45
- Training Time: 24 hours
46
- πŸ“Œ License
47
- This model is licensed under the Apache 2.0 License.
48
-
49
- ---
50
-
51
- ## βœ… **2️⃣ Model Card μž‘μ„± ν›„ μ—…λ‘œλ“œ 방법**
52
- Model CardλŠ” `README.md` νŒŒμΌμ„ Hugging Face μ €μž₯μ†Œμ— μ—…λ‘œλ“œν•˜λ©΄ μžλ™μœΌλ‘œ 반영됨.
53
-
54
- ### **πŸ”Ή 방법 1: Python μ½”λ“œλ‘œ μ—…λ‘œλ“œ**
55
- ```python
56
- from huggingface_hub import HfApi
57
-
58
- api = HfApi()
59
- repo_id = "aripos1/gorani-lora-merged"
60
-
61
- # README.md 파일 μ—…λ‘œλ“œ
62
- api.upload_file(
63
- path_or_fileobj="README.md",
64
- path_in_repo="README.md",
65
- repo_id=repo_id
66
- )
67
-
68
- print("βœ… Model Card (README.md) μ—…λ‘œλ“œ μ™„λ£Œ!")
 
4
  - aripos1/gorani_dataset
5
  language:
6
  - ko
7
+ - en
8
+ - ja
9
  base_model:
10
  - unsloth/Llama-3.2-3B-Instruct-bnb-4bit
11
  pipeline_tag: text-generation
12
  library_name: transformers
13
  ---
14
+ # Gorani Model Card
15
+
16
+ ## μ†Œκ°œ (Introduce)
17
+ 이 λͺ¨λΈμ€ λ²ˆμ—­μ„ μœ„ν•œ λͺ¨λΈμž…λ‹ˆλ‹€. ν•œκ΅­ κ³ μœ μ–΄μ˜ μ •ν™•ν•œ λ²ˆμ—­μ„ μƒμ„±ν•˜κΈ° μœ„ν•΄ ν•œκ΅­μ–΄, μ˜μ–΄, μΌλ³Έμ–΄μ˜ μ–Έμ–΄ 데이터λ₯Ό ν˜Όν•©ν•˜μ—¬ **unsloth/Llama-3.2-1B-Instruct-bnb-4bit**을 ν•™μŠ΅μ‹œμΌœ μƒμ„±λœ **gorani-1B** μž…λ‹ˆλ‹€.
18
+ goraniλŠ” ν˜„μž¬ **ν•œκ΅­μ–΄, μ˜μ–΄, 일본어**만 λ²ˆμ—­μ„ μ§€μ›ν•©λ‹ˆλ‹€.
19
+
20
+ ### λͺ¨λΈ 정보
21
+ - **개발자**: haeun0420
22
+ - **λͺ¨λΈ μœ ν˜•**: **llama**λ₯Ό 기반으둜 ν•˜λŠ” **1B** λ§€κ°œλ³€μˆ˜ λͺ¨λΈμΈ **gorani-1B**
23
+ - **지원 μ–Έμ–΄**: ν•œκ΅­μ–΄, μ˜μ–΄, 일본어
24
+ - **λΌμ΄μ„ΌμŠ€**: **llama**
25
+
26
+ ## Training Hyperparameters
27
+ - **per_device_train_batch_size**: 8
28
+ - **gradient_accumulation_steps**: 1
29
+ - **warmup_steps**: 5
30
+ - **learning_rate**: 2e-4
31
+ - **fp16**: `not is_bfloat16_supported()`
32
+ - **num_train_epochs**: 3
33
+ - **weight_decay**: 0.01
34
+ - **lr_scheduler_type**: "linear"
35
+
36
+ ## ν•™μŠ΅ 데이터
37
+ [데이터셋 링크](https://huggingface.co/datasets/aripos1/gorani_dataset)
38
+
39
+ ## ν•™μŠ΅ μ„±λŠ₯ 비ꡐ
40
+ ![점수 비ꡐ κ·Έλž˜ν”„]
41
+
42
+ ## Training Results
43
+ ![Training Loss κ·Έλž˜ν”„- μŠ€νƒ­μˆ˜μ™€ ν•™μŠ΅ μ§€ν‘œ κ·Έλž˜ν”„]