---
language:
- th
- en
pipeline_tag: text-generation
license: llama3
---
**Llama-3-Typhoon-1.5X-8B-instruct: Thai Large Language Model (Instruct)**

**Llama-3-Typhoon-1.5X-8B-instruct** is an 8 billion parameter instruct model designed for Thai 🇹🇭 language. It demonstrates competitive performance with GPT-3.5-turbo, and is optimized for **application** use cases, **Retrieval-Augmented Generation (RAG), constrained generation**, and **reasoning** tasks.

Built on Typhoon 1.5 8B and Llama 3 8B Instruct. This model is a result of our experiment on **cross-lingual transfer**. It utilizes the [task-arithmetic model editing](https://arxiv.org/abs/2212.04089) technique, combining the Thai understanding capability of Typhoon with the human alignment performance of Llama 3 Instruct.

Remark: To acknowledge Meta's efforts in creating the foundation model and comply with the license, we explicitly include "llama-3" in the model name.

## **Model Description**

- **Model type**: An 8B instruct decoder-only model based on the Llama architecture.
- **Requirement**: Transformers 4.38.0 or newer.
- **Primary Language(s)**: Thai 🇹🇭 and English 🇬🇧
- **License**: [**Llama 3 Community License**](https://llama.meta.com/llama3/license/)

## **Performance**

We evaluated the model's performance in **Language & Knowledge Capabilities** and **Instruction Following Capabilities**.

- **Language & Knowledge Capabilities**:
    - Assessed using multiple-choice question-answering datasets such as ThaiExam and MMLU.
- **Instruction Following Capabilities**:
    - Evaluated based on our beta users' feedback, focusing on two factors:
        - **Human Alignment & Reasoning**: Ability to generate responses that are clear and logically structured across multiple steps.
            - Evaluated using [MT-Bench](https://arxiv.org/abs/2306.05685) — How LLMs can answer embedded knowledge to align with human needs.
        - **Instruction-following**: Ability to adhere to specified constraints in the instruction
            - Evaluated using [IFEval](https://arxiv.org/abs/2311.07911) — How LLMs can follow specified constraints, such as formatting and brevity.

Remark: We developed the TH pair by translating the original datasets into Thai and conducting a human verification on them.

### ThaiExam

| Model | ONET | IC | TGAT | TPAT-1 | A-Level | Average (ThaiExam) | MMLU |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Typhoon-1.5 8B | 0.446 | **0.431** | **0.722** | **0.526** | 0.407 | **0.5028** | 0.6136 |
| Typhoon-1.5X 8B | **0.478** | 0.379 | **0.722** | 0.5 | **0.435** | **0.5028** | 0.6369 |
| gpt-3.5-turbo-0125 | 0.358 | 0.279 | 0.678 | 0.345 | 0.318 | 0.3956 | **0.700**** |

** We report the MMLU score that is reported in GPT-4 Tech Report.

### MT-Bench

| Model | MT-Bench Thai | MT-Bench English |
| --- | --- | --- |
| Typhoon-1.5 8B | 6.402 | 7.275 |
| Typhoon-1.5X 8B | **6.902** | 7.9 |
| gpt-3.5-turbo-0125 | 6.186 | **8.181** |

### IFEval

| Model | IFEval Thai | IFEval English |
| --- | --- | --- |
| Typhoon-1.5 8B | **0.548** | 0.676 |
| Typhoon-1.5X 8B | **0.548** | **0.691** |
| gpt-3.5-turbo-0125 | 0.479 | 0.659 |

## Insight

We utilized **model editing** techniques and found that the most critical feature for generating accurate Thai answers is located in the backend (the upper layers of the transformer block). Accordingly, we incorporated a high ratio of Typhoon components in these backend layers to enhance our model’s performance.

## **Usage Example**

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "scb10x/llama-3-typhoon-v1.5x-8b-instruct"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [...] # add message here

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = model.generate(
    input_ids,
    max_new_tokens=512,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.4,
    top_p=0.95,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))
```

## **Chat Template**

We use the Llama 3 chat template.

```python
{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}
```

## **Intended Uses & Limitations**

This model is experimental and might not be fully evaluated for all use cases. Developers should assess risks in the context of their specific applications.

## **Follow us**

[**https://twitter.com/opentyphoon**](https://twitter.com/opentyphoon)

## **Support**

[**https://discord.gg/CqyBscMFpg**](https://discord.gg/CqyBscMFpg)

## **SCB 10X Typhoon Team**

- Kunat Pipatanakul, Potsawee Manakul, Sittipong Sripaisarnmongkol, Natapong Nitarach, Pathomporn Chokchainant, Kasima Tharnpipitchai
- If you find Typhoon-1.5X useful for your work, please cite it using:

```
@article{pipatanakul2023typhoon,
    title={Typhoon: Thai Large Language Models}, 
    author={Kunat Pipatanakul and Phatrasek Jirabovonvisut and Potsawee Manakul and Sittipong Sripaisarnmongkol and Ruangsak Patomwong and Pathomporn Chokchainant and Kasima Tharnpipitchai},
    year={2023},
    journal={arXiv preprint arXiv:2312.13951},
    url={https://arxiv.org/abs/2312.13951}
}
```

## **Contact Us**

- General & Collaboration: [**kasima@scb10x.com**](mailto:kasima@scb10x.com), [**pathomporn@scb10x.com**](mailto:pathomporn@scb10x.com)
- Technical: [**kunat@scb10x.com**](mailto:kunat@scb10x.com)