This is a continue-pretrained version of Tinyllama-v1.1 tailored for traditional Chinese. The continue-pretraining dataset contains over 10B tokens. Using bfloat16, the VRAM required during inference is only around 3GB!!!

Usage

This is a causal language model not a chat model ! It is not designed to generate human-like responses. It is designed to generate text based on previous text.

python

from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig
import torch
from transformers import TextStreamer

def generate_response(input):
    '''
    simple test for the model
    '''
    # tokenzize the input
    tokenized_input = tokenizer.encode_plus(input, return_tensors='pt').to(device)
    print(tokenized_input['input_ids'])
    # generate the response
    _ = model.generate(
        input_ids=tokenized_input['input_ids'], 
        attention_mask=tokenized_input['attention_mask'],
        pad_token_id=tokenizer.pad_token_id,
        do_sample=True,
        repetition_penalty=1.0,
        max_length=2048,
        streamer=streamer,
    )


if __name__ == '__main__':
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    model = AutoModelForCausalLM.from_pretrained("benchang1110/Taiwan-tinyllama-v1.1-base",attn_implementation="flash_attention_2",device_map=device,torch_dtype=torch.bfloat16)
    tokenizer = AutoTokenizer.from_pretrained("benchang1110/Taiwan-tinyllama-v1.1-base",use_fast=True)
    streamer = TextStreamer(tokenizer)
    while(True):
        text = input("input a simple prompt:")
        generate_response(text)

Training Procedure

The following training hyperparameters are used:

Data size	Global Batch Size	Learning Rate	Epochs	Max Length	Weight Decay
10B	32	5e-5	1	2048	1e-4

Compute Infrastructure

1xA100(80GB), took approximately 200 GPU hours.