mistral-2b-base

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

mistral-2b-base

Welcome to my model card!

This Model feature is ...

• trained by japanese
• trained in two stages: patch level and token level
• Suppression of unknown word generation by using byte fallback in SentencePiece tokenizer and conversion to huggingface Tokenizers format
• Use of Mistral 2B

Yukkuri shite ittene!

<!-- ## Intended uses & limitations More information needed -->

How to use the model

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_path = "ce-lery/mistral-2b-base"
torch.set_float32_matmul_precision('high')

device = "cuda"
if (device != "cuda" and device != "cpu"):
    device = "cpu"

tokenizer = AutoTokenizer.from_pretrained(model_path,use_fast=False)
model = AutoModelForCausalLM.from_pretrained(model_path,
                                             trust_remote_code=True,
                                             ).to(device)

prompt = "自然言語処理とは、"
inputs = tokenizer(prompt,
                   add_special_tokens=True,
                   return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model.generate(
        inputs["input_ids"],
        max_new_tokens=4096,
        do_sample=True,
        early_stopping=False,
        top_p=0.95,
        top_k=50,
        temperature=0.7,
        no_repeat_ngram_size=2,
        num_beams=3
    )

print(outputs.tolist()[0])
outputs_txt = tokenizer.decode(outputs[0])
print(outputs_txt)

Training and evaluation data

40B token. The contents are following.

• Wikipedia
• Wikibooks
• Wikiversity
• CC-100
• OSCAR2109
• mC4 (head 150GB)

Training procedure

Please refer ce-lery/mistral-2b-recipe.
The Guide for this repository is published here. It is written in Japanese.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 128
• total_train_batch_size: 256
• optimizer: Use adamw_bnb_8bit with betas=(0.9,0.95) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine_with_min_lr
• lr_scheduler_warmup_steps: 1000
• num_epochs: 1.0

Training results

Please refer here.

Framework versions

• Transformers 4.46.2
• Pytorch 2.4.0a0+f70bd71a48.nv24.06
• Datasets 2.20.0
• Tokenizers 0.20.3