Mallam 1.1B 4096

🚀 MaLLaM 🌙 1.1B (Malaysia Large Language Model)

MaLLaM 1.1B is a large language model specifically pretrained on Malaysian text. It uses the Mistral architecture and has been trained from scratch with 1.1B parameters on 90B Malaysian text tokens, offering a context length of 4096.

🚀 Quick Start

Prerequisites

Before using MaLLaM 1.1B, make sure you have the necessary libraries installed. You can use the following code to load the model and tokenizer:

from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
import torch

TORCH_DTYPE = 'bfloat16'
nf4_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type='nf4',
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=getattr(torch, TORCH_DTYPE)
)

tokenizer = AutoTokenizer.from_pretrained('mesolitica/mallam-1.1B-4096')
model = AutoModelForCausalLM.from_pretrained(
    'mesolitica/mallam-1.1B-4096',
    use_flash_attention_2 = True,
    quantization_config = nf4_config
)
prompt = '<s>nama saya'
inputs = tokenizer([prompt], return_tensors='pt', add_special_tokens=False).to('cuda')

generate_kwargs = dict(
    inputs,
    max_new_tokens=512,
    top_p=0.95,
    top_k=50,
    temperature=0.9,
    do_sample=True,
    num_beams=1,
    repetition_penalty=1.05,
)
r = model.generate(**generate_kwargs)

✨ Features

• Pretrained on Malaysian Text: MaLLaM 1.1B is trained on 90B Malaysian text tokens, which makes it well - adapted to the Malaysian language context.
• Mistral Architecture: Utilizes the Mistral architecture, which provides efficient and powerful language processing capabilities.
• 4096 Context Length: Supports a context length of 4096, allowing for more comprehensive text understanding and generation.

📦 Installation

There is no specific installation process described in the original document. Usually, you need to have the transformers library installed. You can install it using pip install transformers.

📚 Documentation

• README: Check the README at https://github.com/mesolitica/malaya/tree/5.1/pretrained-model/mistral.
• Technical Report: Refer to the technical report at https://github.com/mesolitica/malaya/wiki/MaLLaM-%F0%9F%8C%99-Malaysia-Large-Language-Model.

🔧 Technical Details

• Training Data: The model is trained on 90B tokens, which are gathered at https://github.com/malaysia-ai/dedup-text-dataset/tree/main/pretrain-llm.
• Training Infrastructure: A Ray cluster is used to train the model on 5 nodes of 4x A100 80GB. You can find more details at https://github.com/malaysia-ai/jupyter-gpu/tree/main/ray.

📄 Additional Resources

• WandB: Monitor the training process on https://wandb.ai/mesolitica/pretrain-mistral-1.1b?workspace=user-husein-mesolitica.
• WandB Report: Read the WandB report at https://wandb.ai/mesolitica/pretrain-mistral-3b/reports/Pretrain-Larger-Malaysian-Mistral--Vmlldzo2MDkyOTgz.