japanese-gpt-neox-3.6b-instruction-sft-v2

Overview

This repository provides a Japanese GPT-NeoX model of 3.6 billion parameters. The model is based on rinna/japanese-gpt-neox-3.6b and has been finetuned to serve as an instruction-following conversational agent.

This model slightly differs from the previous SFT model rinna/japanese-gpt-neox-3.6b-instruction-sft , where a different data split is used for training.

• Model architecture

  A 36-layer, 2816-hidden-size transformer-based language model.

• SFT vs. previous SFT evaluation

  We conducted ChatGPT-based automated evaluation on 100 prompts to assess the performance difference between this SFT model and the previous SFT model.

  this SFT vs. previous SFT	win	tie	loss
  ChatGPT auto. evaluation	55 %	0%	45%

• Finetuning

  The finetuning data is the subset of the following datasets and has been translated into Japanese.

  • Anthropic HH RLHF data
  • FLAN Instruction Tuning data
  • Stanford Human Preferences Dataset

  The data will not be released.

• Model Series

  Variant	Link
  3.6B PPO	https://huggingface.co/rinna/japanese-gpt-neox-3.6b-instruction-ppo
  3.6B SFT-v2	https://huggingface.co/rinna/japanese-gpt-neox-3.6b-instruction-sft-v2
  3.6B SFT	https://huggingface.co/rinna/japanese-gpt-neox-3.6b-instruction-sft
  3.6B pretrained	https://huggingface.co/rinna/japanese-gpt-neox-3.6b

• Authors

  Tianyu Zhao and Kei Sawada

I/O Format

A special format has been adopted to construct inputs.

• An input prompt is formatted as a conversation between ユーザー and システム .
• Each input utterance consists of (1) its speaker ( "ユーザー" or "システム" ), (2) a colon ( ":" ), (3) a whitespace ( " " ), and (4) utterance text (e.g. "世界で一番高い山は？" ).
• The input prompt should be ended with "システム: " to acknowledge the model to generate a response.
• Since the model's tokenizer does not recognize "\n" , a special newline symbol "<NL>" is used instead.
• All the newlines in input and output utterances should be replaced with "<NL>" .
• All the utterances in the input prompt should be separated by "<NL>" .

Following is an example to construct an input from a conversation.

prompt = [
    {
        "speaker": "ユーザー",
        "text": "コンタクトレンズを慣れるにはどうすればよいですか？"
    },
    {
        "speaker": "システム",
        "text": "これについて具体的に説明していただけますか？何が難しいのでしょうか？"
    },
    {
        "speaker": "ユーザー",
        "text": "目が痛いのです。"
    },
    {
        "speaker": "システム",
        "text": "分かりました、コンタクトレンズをつけると目がかゆくなるということですね。思った以上にレンズを外す必要があるでしょうか？"
    },
    {
        "speaker": "ユーザー",
        "text": "いえ、レンズは外しませんが、目が赤くなるんです。"
    }
]
prompt = [
    f"{uttr['speaker']}: {uttr['text']}"
    for uttr in prompt
]
prompt = "<NL>".join(prompt)
prompt = (
    prompt
    + "<NL>"
    + "システム: "
)
print(prompt)
# "ユーザー: コンタクトレンズを慣れるにはどうすればよいですか？<NL>システム: これについて具体的に説明していただけますか？何が難しいのでしょうか？<NL>ユーザー: 目が痛いのです。<NL>システム: 分かりました、コンタクトレンズをつけると目がかゆくなるということですね。思った以上にレンズを外す必要があるでしょうか？<NL>ユーザー: いえ、レンズは外しませんが、目が赤くなるんです。<NL>システム: "

How to use the model

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("rinna/japanese-gpt-neox-3.6b-instruction-sft-v2", use_fast=False)
model = AutoModelForCausalLM.from_pretrained("rinna/japanese-gpt-neox-3.6b-instruction-sft-v2")

if torch.cuda.is_available():
    model = model.to("cuda")

token_ids = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")

with torch.no_grad():
    output_ids = model.generate(
        token_ids.to(model.device),
        do_sample=True,
        max_new_tokens=128,
        temperature=0.7,
        repetition_penalty=1.1,
        pad_token_id=tokenizer.pad_token_id,
        bos_token_id=tokenizer.bos_token_id,
        eos_token_id=tokenizer.eos_token_id
    )

output = tokenizer.decode(output_ids.tolist()[0][token_ids.size(1):])
output = output.replace("<NL>", "\n")
print(output)
"""わかりました。まずは、コンタクトレンズを長時間着用することによる目の乾燥を防ぐことができます。また、毎日同じ時間帯にコンタクトレンズを着用してみることもできます。そして、コンタクトレンズが目に合わないような場合は、新しいものを試してみる必要があります。</s>"""

Tokenization

The model uses a sentencepiece -based tokenizer.

• The tokenizer has a vocabulary size of 32,000.
• It uses sentencepiece's byte fallback feature to decompose unknown text pieces into UTF-8 byte pieces and to avoid producing <UNK> tokens.
• sentencepiece's --add_dummy_prefix option was turned off so that a leading whitespace will not be prepended automatically.

    print(tokenizer.tokenize("吾輩は猫である"))
    # ['吾', '輩', 'は', '猫', 'である']
    # instead of ['▁', '吾', '輩', 'は', '猫', 'である'] as in rinna/japanese-gpt-1b
  

• sentencepiece's --remove_extra_whitespaces option was turned off so that leading, trailing, and duplicate whitespaces are reserved.

    print(tokenizer.tokenize("  吾輩は  猫である   "))
    # ['▁', '▁', '吾', '輩', 'は', '▁', '▁', '猫', 'である', '▁', '▁', '▁']
    # instead of ['▁', '吾', '輩', 'は', '▁猫', 'である'] as in rinna/japanese-gpt-1b
  

• Don't forget to set use_fast=False to make the above features function correctly.

    good_tokenizer = AutoTokenizer.from_pretrained("rinna/japanese-gpt-neox-3.6b", use_fast=False)
    bad_tokenizer = AutoTokenizer.from_pretrained("rinna/japanese-gpt-neox-3.6b")
  
    print(good_tokenizer.decode(good_tokenizer.encode("გამარჯობა  吾輩は  猫である   ")))
    # 'გამარჯობა  吾輩は  猫である   </s>'
    print(bad_tokenizer.decode(bad_tokenizer.encode("გამარჯობა  吾輩は  猫である   ")))
    # 'გამარ[UNK]ობა 吾輩は 猫である </s>'
  

Licenese

The MIT license