---
license: apache-2.0
language:
- vi
- en
---

<p align="center">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/63905e87df447b438817b2cd/QFhLKQlWeyO9XumtyghVo.jpeg" alt="Image" style="width: 400px; height: auto; border-radius: 10px;" />
</p>


## Model Details

- **Developed by:** Tuan Pham (FPTU HCM Student)
  - Contact me at: weekend.2810@gmail.com or tuanpmse160561@fpt.edu.vn
  - Looking for intern opportunity :D 
- **Model type:** Llama2-7B Decoder-only
- **Finetuned from model :**
  * meta-llama/Llama-2-7b
  * bkai-foundation-models/vietnamese-llama2-7b-120GB
  * yeen214/llama2_7b_merge_orcafamily.
- **Bilingual support :** English and Vietnamese

### Model Description

<!-- Provide a longer summary of what this model is. -->

This model is a proof of effort that one man can fine-tune his own model to reach SOTA.

### Model Sources

<!-- Provide the basic links for the model. -->

- **Repository:** 
  * Training: https://github.com/vTuanpham/Vietnamese_QA_System
  * Data: https://github.com/vTuanpham/Large_dataset_translator
- **Paper:** ...
- **Demo:** ...

## Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

### Prompt template

```
[SYSTEM_PROMPT]

 ####### Instruction:
[INPUT]

 %%%%%%% Response:
[RESPONSE]
```
Recommend keeping the system prompt in english.
## How to Get Started with the Model

Use the code below to get started with the model.
```python
import torch
from torch.cuda.amp import autocast
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer, pipeline


def prompt_format(system_prompt, instruction):
    prompt = f"""{system_prompt}

 ####### Instruction:
{instruction}

 %%%%%%% Response:

"""
    return prompt

system_prompt = """
You're an AI Large Language Model developed(created) by an AI developer named Tuấn, the architecture of you is decoder-based LM, your task are to think loudly step by step before give a good and relevant response
to the user request, answer in the language the user preferred.

The AI has been trained to answer questions, provide recommendations, and help with decision making. The AI thinks outside the box and follows the user requests
"""
instruction = "Xin chào"

formatted_prompt = prompt_format(system_prompt, instruction)
print(formatted_prompt)


model_name = "1TuanPham/T-Llama"
model = AutoModelForCausalLM.from_pretrained(model_name,
                                             torch_dtype=torch.bfloat16,
                                             use_cache=True,
                                             device_map="auto"
                                             )
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
streamer = TextStreamer(tokenizer, skip_special_tokens=True)
pipe = pipeline("text-generation", model=base_model, tokenizer=tokenizer, streamer=streamer)

with autocast():
  output_default = pipe(formatted_prompt, pad_token_id=50256, max_new_tokens=128)

```
Example output:
```bash
Xin chào! Tôi là một AI được phát triển bởi một AI nhà phát triển tên là Tuấn. Tôi được thiết kế để giúp đỡ người dùng bằng cách trả lời các câu hỏi, đưa ra đề xuất và hỗ trợ trong quá trình ra quyết định.
Tôi có thể hỗ trợ bạn bằng cách nghĩ ra các câu trả lời hay và phù hợp cho các câu hỏi của bạn.
```

Note: 120GB of pre-trained Vietnamese  data might not be enough for a general question about Vietnamese events.

Here is a kaggle script to quickly test the model:
* https://www.kaggle.com/code/tuanphamm/t-llama-test

## Training Details

**Hardware Type:**
  * GPU: VGA NVIDIA Tesla P100 16GB
  * SYSTEM RAM: 29GB
  
**Hours used:** ~47.5 days Approx*

### Training Data

* BactrianX 
* OpenOrca_translated 
* WizardLM_70k_translated 
* TigerLabMathInstruct_translated_vi 
* GradeSchoolMathInstruct_translated 
* vilm_lima-vi
* MTEngVietnamese 
* databricks_dolly15k_translated 
* AlpacaCleaned_translated 
* databricks_dolly15k
* OpenOrca
* GradeSchoolMathInstruct 
* AlpacaCleaned
* WebglmQA

### Training Procedure

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->

* Learning rate: 2e-5 cosine
* Optimizer: PagedLion8bit
* QLora: rank: 64 /Q: 4-bit
  
  - 250k examples of 70% Vietnamese 30% English for 3.37 epoch
  - 350k examples of 60% Vietnamese 40% English for 1.4 epoch

### Training loss

![image/png](https://cdn-uploads.huggingface.co/production/uploads/63905e87df447b438817b2cd/rV8Go_YFZv7QcR_FhFxp-.png)

Each line is 12 hours
## Evaluation

<!-- This section describes the evaluation protocols and provides the results. -->

![image/png](https://cdn-uploads.huggingface.co/production/uploads/63905e87df447b438817b2cd/z1ZTm7Tab4tQbVPgQW1hU.png)

Our model currently sits at TOP-5 on the VMLU benchmark

## Citation

<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
```bibtex
@online{t-llama,
  author = {Pham Minh Tuan},
  title = {T-Llama: A New Language Model for Vietnamese},
  year = 2024,
  url = {https://github.com/vTuanpham/Vietnamese_QA_System}
}
```