MaLLaM 🌙 3B (Malaysia Large Language Model), Pretrain 3B 4096 context length on Malaysian text

Pretrain from scratch 3B parameters using Mistral architecture on 90B Malaysian text tokens.

README at https://github.com/mesolitica/malaya/tree/5.1/pretrained-model/mistral

• Trained on 90B tokens, gathered at https://github.com/malaysia-ai/dedup-text-dataset/tree/main/pretrain-llm
• We use Ray cluster to train on 5 nodes of 4x A100 80GB, https://github.com/malaysia-ai/jupyter-gpu/tree/main/ray

WandB, https://wandb.ai/mesolitica/pretrain-mistral-3b?workspace=user-husein-mesolitica

WandB report, https://wandb.ai/mesolitica/pretrain-mistral-3b/reports/Pretrain-Larger-Malaysian-Mistral--Vmlldzo2MDkyOTgz

Technical report, https://github.com/mesolitica/malaya/wiki/MaLLaM-%F0%9F%8C%99-Malaysia-Large-Language-Model

how-to

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-3B-4096')
model = AutoModelForCausalLM.from_pretrained(
    'mesolitica/mallam-3B-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)