This repository contains fine-tuning scripts for both supervised fine-tuning (SFT) and alignment scripts. Our goal is to create a model-agnostic fine-tuning pipeline and evaluation scripts focusing on the usability of the Thai language. The repository consists of three training scripts: (i) supervised fine-tuning (SFT), (ii) direct preference optimization (DPO), and (iii) odds ratio preference optimization (ORPO).

Content

Supported base LLMs
Released Models
Evaluation
Installation
Prepare Dataset (Optional)
Fine-tuning
Inference
Deployment
Retrieval Augmented Generation (RAG)
Acknowledgements
Future Plans
Citation

💡 Supported base LLMs

Here is the list of supported base LLMs that we have tested on our scripts.

LLaMa3
SeaLLMs
PolyLM
Typhoon
SEA-LION (Please refer to GitHub: vistec-AI/WangchanLion for the full detail)
Gemma 2

🤖 Released Models

We apply our fine-tuning pipeline to various open-source models and publish their weights as follows:

Demo models

The models that trained on small instruction datasets

LLaMa3-8b-WangchanX-sft-Demo
PolyLM-13b-WangchanX-sft-Demo
typhoon-7b-WangchanX-sft-Demo

Full models

The models that trained on large instruction datasets. For reproducibility, we provide the scripts for dataset collection and preprocessing in this repository.

LLaMa3-8b-WangchanX-sft
SeaLion-7b-WangchanX-sft

⚡ Evaluation

We evaluate LLMs using the Benchmark Suite for Southeast Asian Languages. For detailed information on our evaluation methodology and benchmarking process, visit the SEACrowd project repository.

NLU

NLG

📦 Installation

Please install all dependencies in requirements.txt using pip install as

pip3 install -r requirements.txt

Please install Flash Attention 2 using pip install as

pip3 install flash-attn --no-build-isolation

Go to the Fine-tuning section and select the training strategy that is suitable for your constraints.

📋 Prepare Dataset (Optional)

Using a Custom Demo Dataset

If you want to use a custom dataset, you need to reformat the file by editing it.

python3 reformat.py

If you want to use the demo dataset, you can download it from this.

This dataset includes 6 datasets:

pythainlp/han-instruct-dataset-v2.0
databricks/databricks-dolly-15k
databricks/databricks-dolly-15k (translated English to Thai by Gemini)
math_14k
math_14k (translated English to Thai by Gemini)
iapp_wiki_qa_squad

Using the Full Dataset

Creating the Dataset:
- Go to the create dataset script page.
- Download the script provided there.
- Run the following command in your terminal:
```
python main.py --output_dir /<path>/flan_dataset
```
  This will create the full dataset in a directory called flan_dataset.

Updating the Configuration:

Find the configuration file for your specific model and training mode.
The file will be located at: recipes/<model_name>/<mode>/config_<method>.yaml
For example, if you're using the LLaMA3-8b model for supervised fine-tuning (sft), the file would be: recipe/llama3-8b/sft/config_full.yaml

Open this file and update the dataset_mixer section to point to your newly created dataset:

# Data training arguments
chat_template: "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ '<|user|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'system' %}\n{{ '<|system|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'assistant' %}\n{{ '<|assistant|>\n'  + message['content'] + eos_token }}\n{% endif %}\n{% if loop.last and add_generation_prompt %}\n{{ '<|assistant|>' }}\n{% endif %}\n{% endfor %}"
dataset_mixer:
  /<path>/flan_dataset: 1.0 # <- This is the path to your newly created dataset
dataset_splits:
  - train
preprocessing_num_workers: 12

The key change is in the dataset_mixer section, where /<path>/flan_dataset should be the path to your created dataset.

By following these steps, you'll have prepared the full dataset and updated your configuration file to use it for training your model.

🛠 Fine-tuning

To start fine-tuning your own LLM, we recommend using QLoRa fine-tuning because it consumes much fewer resources compared to fully fine-tuning the LLM. Please note that the provided examples are all LLaMa3. The main template for the script is structured as

{RUNNER} scripts/run_{MODE}.py {RECIPE}

The main parameters are

Parameter	Description
`RUNNER`	Can be `python` for single-GPU fine-tuning or `accelerate` with the argument `--config_file {ACCELERATION_CONFIG}` for multi-GPU training.
`ACCELERATION_CONFIG`	The mode to launch the trainer in multiple setups. Mainly, there are vanilla multi-GPU and ZeRO3 offloading for lower GPU memory usage with IO overhead. Available configurations are in `recipes/accelerate_configs`.
`MODE`	Can be `sft` (supervised fine-tuning) or `dpo` (direct preference optimization).
`RECIPE`	Based on the model types in the `recipes` folder.

QLoRa fine-tuning example

The simplest way to start fine-tuning your LLM is to use plain Python on a single GPU. You can do the supervised fine-tuning (SFT) and direct preference optimization (DPO) as in the following step.

# Step 1 - SFT
python scripts/run_sft.py recipes/llama3-8b/sft/config_qlora.yaml
# Step 2 - DPO (optional)
python scripts/run_dpo.py recipes/llama3-8b/dpo/config_qlora.yaml

Alternatively, you can exploit multi-gpus training by using the bellowing scripts.

# Step 1 - SFT
ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/multi_gpu.yaml --num_processes=4 scripts/run_sft.py recipes/llama3-8b/sft/config_qlora.yaml
# Step 2 - DPO
ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/multi_gpu.yaml --num_processes=4 scripts/run_dpo.py recipes/llama3-8b/dpo/config_qlora.yaml

Please note that the number of arguments num_processes should be the number of your available GPUs. We use the the default num_processes=4.

Full fine-tuning example

You can fine-tune the whole model using the following scripts.

# Step 1 - SFT
ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/multi_gpu.yaml scripts/run_sft.py recipes/llama3-8b/sft/config_full.yaml
# Step 2 - DPO
ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/multi_gpu.yaml scripts/run_dpo.py recipes/llama3-8b/dpo/config_full.yaml

In case you have limited GPU resources but still want to do the full fine-tuing, please consider using DeepSpeed ZeRO3. By adding config_file argument, you are good to go!

# Step 1 - SFT
ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/deepspeed_zero3.yaml scripts/run_sft.py recipes/llama3-8b/sft/config_full.yaml
# Step 2 - DPO
ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/deepspeed_zero3.yaml scripts/run_dpo.py recipes/llama3-8b/dpo/config_full.yaml

🌟 Inference

Prepare your model and tokenizer

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# Model path
path = "airesearch/LLaMa3-8b-WangchanX-sft-Full"

# Device
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(path, use_fast=False)
model = AutoModelForCausalLM.from_pretrained(path, device_map="auto")

Define chat messages

messages = [
    {"role": "user", "content": "ลิเก กับ งิ้ว ต่างกันอย่างไร"},
]

Tokenize chat messages

tokenized_chat = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(device)
print(tokenizer.decode(tokenized_chat[0]))

Output:

<|user|>
ลิเก กับ งิ้ว ต่างกันอย่างไร<|end_of_text|>
<|assistant|>

Generate responses

outputs = model.generate(tokenized_chat, max_length=2048)
print(tokenizer.decode(outputs[0]))

Output:

<|user|>
ลิเก กับ งิ้ว ต่างกันอย่างไร<|end_of_text|>
<|assistant|>
ก่อนอื่นเราต้องรู้ความหมายของคำทั้งสอง คำว่า ลิเก เป็นศิลปะการแสดงแบบดั้งเดิมในประเทศไทย ส่วนคำว่า งิ้วน่าจะเป็นการนำภาษาไทยมาแปลจากคำว่า อินโดปีเลีย (indoplea) ซึ่งเป็นชื่อเรียกดนตรีที่มีต้นกำเนิดจากรัฐอุตตาร์ประเทศ ในอินเดีย และได้แพร่หลายไปยังเอเชียตะวันออกเฉียงใต้ โดยเฉพาะสาธารณรัฐประชาชนจีนและเวียดนาม จึงทำให้เกิดคำว่า งิ้วด้วย แต่ทุกคนไม่รู้ว่ามันก็คืออะไรจริง ๆ แล้ว มันมีความแตกต่างกันมาก เพราะถ้าไปถามชาวบ้านบางแห่งอาจจะบอกว่าเป็นอีกประเภทหนึ่งของเพลงโบราณหรือเพลงพื้นเมือง หรือถ้าพูดตามหลักทางประวัติศาสตร์ก็จะกล่าวว่านั่นคือ การขับร้องเพลงที่ใช้รูปแบบการประสานเสียงแบบฮินดู-ซิกห์วัล ที่ผสมผสานระหว่างภาษาอังกฤษ ภาษาจีนกลาง ภาษาพม่า และภาษาทางเหนือกับภาษาลาว รวมถึงภาษากลุ่มออสเตรโลไนว์ในอดีต ดังนั้นตอนนี้คุณสามารถสรุปได้อย่างแม่นยำว่าสองอย่างเหล่านี้แตกต่างกันอย่างไร: ลิเก คือ ศิลปะการแสดงที่มีมายาวนานกว่า 100 ปีในประเทศไทย เช่น ลิเกล้านนา, ลิเกตลุง, ลิเกล้อ ฯลฯ ขณะที่ งิ้ว หมายถึง เพลงประสานเสียงที่มีรากเหง้าของวงการเพลงคลาสสิคในอินเดีย และแพร่กระจายในเอเชียตะวันตกเฉียงใต้เป็นสิ่งแรกๆ หลังจากการเผยแผ่ศาสนายุคแรกๆ นอกจากนี้ ยังมีการรวมแนวเพลงเพื่อรวมเข้ากับการเต้นร่วมสมัยและบทละครที่มีอิทธิพลจากวรรณกรรมจีน<|end_of_text|>

🚀 Deployment

See Deployments.md for details on deploying pre-trained Large Language Models (LLMs) using Text Generation Inference (TGI), LocalAI, and Ollama frameworks.

✨ Retrieval Augmented Generation (RAG)

See RAG.md for details on setting up a Retrieval Augmented Generation system using Flowise, LocalAI, and Ollama frameworks for enhancing language model generation with retrieved knowledge.

🙏 Acknowledgements

We would like to thank all codes and structures from alignment-handbook. This project is sponsored by VISTEC, PTT, SCBX, and SCB.

📅 Future Plans

Here are some future plans and what we are doing:

Adding model and codes for ORPO. Currently, we have codes and preliminary models from the ORPO technique. We are planning to release them soon.
Thai LLMs benchmark. We are planning to create a machine reading comprehension leaderboard for Thai LLMs. We are happy for any ideas or contributions from everyone.

📜 Citation

If you use WangchanX or WangchanX Eval in your project or publication, please cite the library as follows

@misc{phatthiyaphaibun2024wangchanlion,
      title={WangchanLion and WangchanX MRC Eval},
      author={Wannaphong Phatthiyaphaibun and Surapon Nonesung and Patomporn Payoungkhamdee and Peerat Limkonchotiwat and Can Udomcharoenchaikit and Jitkapat Sawatphol and Chompakorn Chaksangchaichot and Ekapol Chuangsuwanich and Sarana Nutanong},
      year={2024},
      eprint={2403.16127},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

README.md

README.md

Content

💡 Supported base LLMs

🤖 Released Models

Demo models

Full models

⚡ Evaluation

NLU

NLG

📦 Installation

📋 Prepare Dataset (Optional)

Using a Custom Demo Dataset

Using the Full Dataset

🛠 Fine-tuning

🌟 Inference

Prepare your model and tokenizer

Define chat messages

Tokenize chat messages

Generate responses

🚀 Deployment

✨ Retrieval Augmented Generation (RAG)

🙏 Acknowledgements

📅 Future Plans

📜 Citation

Files

README.md

Latest commit

History

README.md

File metadata and controls

Content

💡 Supported base LLMs

🤖 Released Models

Demo models

Full models

⚡ Evaluation

NLU

NLG

📦 Installation

📋 Prepare Dataset (Optional)

Using a Custom Demo Dataset

Using the Full Dataset

🛠 Fine-tuning

🌟 Inference

Prepare your model and tokenizer

Define chat messages

Tokenize chat messages

Generate responses

🚀 Deployment

✨ Retrieval Augmented Generation (RAG)

🙏 Acknowledgements

📅 Future Plans

📜 Citation