myBit-Llama2-jp-127M-4

This model has 127M parameters.
The model is a pre-trained Bit-Llama2 of Parameters with only 1 epoch on a Japanese dataset. The dataset used is range3/wiki40b-ja.

• Loss: 2.9790

Model description

Github：　BitNet-b158
More information about this model can be found in the following pages:

• BitNet&BitNet b158の実装①
• BitNet&BitNet b158の実装②

How to use

1. install the library

!pip install mybitnet==0.2.3
!pip install -U accelerate transformers==4.38.2
!pip install torch

2. get model

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "HachiML/myBit-Llama2-jp-127M-4"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)
print(model)

3. inference

prompt = "昔々あるところに、"
input_ids = tokenizer.encode(
    prompt,
    return_tensors="pt"
)
tokens = model.generate(
    input_ids.to(device=model.device),
    max_new_tokens=128,
)

out = tokenizer.decode(tokens[0], skip_special_tokens=True)
print(out)

Intended uses & limitations

More information needed

Training and evaluation data

• range3/wiki40b-ja

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0024
• train_batch_size: 96
• eval_batch_size: 96
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: polynomial
• lr_scheduler_warmup_steps: 5000
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
4.8696	0.05	2000	3.8588
3.7027	0.1	4000	3.6106
3.5648	0.15	6000	3.5014
3.448	0.2	8000	3.4153
3.3884	0.25	10000	3.3650
3.3462	0.29	12000	3.3280
3.3155	0.34	14000	3.3053
3.2932	0.39	16000	3.2891
3.2762	0.44	18000	3.2673
3.2594	0.49	20000	3.2533
3.2432	0.54	22000	3.2398
3.2286	0.59	24000	3.2186
3.2083	0.64	26000	3.1957
3.1867	0.69	28000	3.1769
3.1676	0.74	30000	3.1568
3.14	0.79	32000	3.1286
3.114	0.83	34000	3.1006
3.0848	0.88	36000	3.0696
3.0511	0.93	38000	3.0301
3.005	0.98	40000	2.9790

Framework versions

• Transformers 4.38.2
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2