README.md

---
library_name: transformers
license: apache-2.0
datasets:
- nampdn-ai/tiny-codes
- nlpai-lab/openassistant-guanaco-ko
- philschmid/guanaco-sharegpt-style
language:
- ko
- en
inference: false
tags:
- unsloth
- phi-3
pipeline_tag: text-generation
---

# Phi-3-medium-4k-instruct-ko-poc-v0.1

## Model Details
This model is trained using unsloth toolkit based on Microsoft's phi-3 Phi-3-medium-4k-instruct model (https://huggingface.co/unsloth/Phi-3-medium-4k-instruct) with some Korean instruction data added to enhance its Korean generation performance

Since my role is not as a working developer, but as ML Technical Specialist helping customers with quick PoCs/prototypes, and I was limited by Azure GPU resources available, I only trained with 40,000 samples on a single VM Azure Standard_NC24ads_A100_v4 for PoC purposes. Because I have not done any tokenizer extensions, you need a lot more tokens than English for text generation.

### Dataset

The dataset used for training is as follows. To prevent catastrophic forgetting, I included non-Korean corpus as training data. Note that we did not use all of the data, but only sampled some of it. Korean textbooks were converted to Q&A format. The Guanaco dataset has been reformatted to fit the multiturn format like <|user|>\n{Q1}<|end|>\n<|assistant|>\n{A1}<|end|>\n<|user|>\n{Q2}<|end|>\n<|assistant|>\n{A2}<|end|>.

- Korean textbooks (https://huggingface.co/datasets/nampdn-ai/tiny-codes)
- Korean translation of Guanaco (https://huggingface.co/datasets/nlpai-lab/openassistant-guanaco-ko)
- Guanaco Sharegpt style (https://huggingface.co/datasets/philschmid/guanaco-sharegpt-style)


## How to Get Started with the Model

### Code snippets
```python
### Load model
import torch
from unsloth import FastLanguageModel
from unsloth.chat_templates import get_chat_template
from transformers import TextStreamer

max_seq_length = 2048 # Choose any! We auto support RoPE Scaling internally!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
model_path = "daekeun-ml/Phi-3-medium-4k-instruct-ko-poc-v0.1"

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = model_tar_dir, # Choose ANY! eg teknium/OpenHermes-2.5-Mistral-7B
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
    # token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
)
tokenizer = get_chat_template(
    tokenizer,
    chat_template = "phi-3", # Supports zephyr, chatml, mistral, llama, alpaca, vicuna, vicuna_old, unsloth
    mapping = {"role" : "from", "content" : "value", "user" : "human", "assistant" : "gpt"}, # ShareGPT style
)

params = {
    "max_new_tokens": 256,
    "use_cache": True,
    "temperature": 0.05,
    "do_sample": True
}

### Inference
FastLanguageModel.for_inference(model) # Enable native 2x faster inference

# 1st example
messages = [
    {"from": "human", "value": "Continue the fibonnaci sequence in Korean: 1, 1, 2, 3, 5, 8,"},  
    {"from": "assistant", "value": "피보나치 수열의 다음 숫자는 13, 21, 34, 55, 89 등입니다. 각 숫자는 앞의 두 숫자의 합입니다."},    
    {"from": "human", "value": "Compute 2x+3=12 in Korean"}, 
]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize = True,
    add_generation_prompt = True, # Must add for generation
    return_tensors = "pt",
).to("cuda")

text_streamer = TextStreamer(tokenizer)
_ = model.generate(input_ids = inputs, streamer = text_streamer, **params)

# 2nd example
messages = [
    {"from": "human", "value": "What is Machine Learning in Korean?"},  
    {"from": "assistant", "value": "인공지능의 한 분야로 방대한 데이터를 분석해 향후 패턴을 예측하는 기법입니다."},    
    {"from": "human", "value": "What is Deep Learning in Korean?"}, 
]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize = True,
    add_generation_prompt = True, # Must add for generation
    return_tensors = "pt",
).to("cuda")

from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer)
_ = model.generate(input_ids = inputs, streamer = text_streamer, **params)
```

### Inference results
```python
# 1st example
<s><|user|> Continue the fibonnaci sequence in Korean: 1, 1, 2, 3, 5, 8,<|end|><|assistant|> 피보나치 수열의 다음 숫자는 13, 21, 34, 55, 89 등입니다. 각 숫자는 앞의 두 숫자의 합입니다.<|end|><|user|> Compute 2x+3=12 in Korean<|end|><|assistant|> 방정식 2x + 3 = 12에서 x를 풀려면 다음 단계를 따르십시오.

1. 방정식의 양쪽에서 3을 빼서 방정식의 한쪽에 대해 x를 분리합니다.

2x + 3 - 3 = 12 - 3

2x = 9

2. 이제 방정식의 양쪽을 2로 나누어 x의 값을 구합니다.

2배 / 2 = 9 / 2

x = 4.5

따라서 방정식 2x + 3 = 12에 대한 해는 x = 4.5입니다.<|end|>

# 2nd example
<s><|user|> What is Machine Learning in Korean?<|end|><|assistant|> 인공지능의 한 분야로 방대한 데이터를 분석해 향후 패턴을 예측하는 기법입니다.<|end|><|user|> What is Deep Learning in Korean?<|end|><|assistant|> 복잡한 데이터 세트를 분석하고 복잡한 패턴을 인식하고 학습하는 데 사용되는 딥러닝은 많은 레이어로 구성된 신경망의 하위 집합입니다. 이 기술은 이미지 인식, 자연어 처리 및 자율 운전과 같은 다양한 응용 분야에서 큰 발전을 이뤘습니다.<|end|>
```

### References
- Base model: [unsloth/Phi-3-medium-4k-instruct](https://huggingface.co/unsloth/Phi-3-medium-4k-instruct)

## Notes 

### License

apache 2.0; The license of phi-3 is MIT, but I considered the licensing of the dataset and library used for training.

### Caution
This model was created as a personal experiment, unrelated to the organization I work for. The model may not operate correctly because separate verification was not performed. Please be careful unless it is for personal experimentation or PoC (Proof of Concept)!