thumbnail: https://github.com/rinnakk/japanese-pretrained-models/blob/master/rinna.png
license: mit
datasets:
- Anthropic/hh-rlhf
language:
- ja
- en
inference: false
bilingual-gpt-neox-4b-instruction-ppo
Overview
This repository provides an English-Japanese bilingual GPT-NeoX model of 3.8 billion parameters.
The model is based on rinna/bilingual-gpt-neox-4b-instruction-sft and has been aligned to serve as an instruction-following conversational agent.
Model architecture
A 36-layer, 2816-hidden-size transformer-based language model.
RLHF
Following the OpenAI InstructGPT paper, Reinforcement Learning from Human Feedback (RLHF) has been applied to aligning the model's behaviour with input instructions. Particularly, the model has been trained in two stages, i.e. Supervised Fine-Tuning (SFT) and PPO-based Reinforcement Learning (RL).
- The first SFT stage produces
rinna/bilingual-gpt-neox-4b-instruction-sft. - The second RL stage produces this model.
- The first SFT stage produces
Reinforcement learning
We used CarperAI/trlx and its implementation of the PPO algorithm for the RL stage.
The RL data is the subset of the following dataset and has been translated into Japanese.
Model Series
Authors
Benchmarking
Our evaluation experiments suggest that the PPO does not particularly improve the model's performance on the Japanese LLM benchmark in comparison with Bilingual GPT-NeoX 4B SFT, but we have seen better conversation experience on the PPO model than its SFT counterpart.
- The 4-task average accuracy is based on results of JCommonsenseQA, JNLI, MARC-ja, and JSQuAD.
- The 6-task average accuracy is based on results of JCommonsenseQA, JNLI, MARC-ja, JSQuAD, XWinograd, and JAQKET-v2.
| Model | 4-task average accuracy | 6-task average accuracy |
|---|---|---|
| bilingual-gpt-neox-4b-instruction-ppo | 61.01 | 61.16 |
| bilingual-gpt-neox-4b-instruction-sft | 61.02 | 61.69 |
| bilingual-gpt-neox-4b | 56.12 | 51.83 |
| japanese-gpt-neox-3.6b-instruction-ppo | 59.86 | 60.07 |
| japanese-gpt-neox-3.6b | 55.07 | 50.32 |
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. - All the utterances in the input prompt should be separated by a newline
\n.
Following is an example to construct input from a conversation.
prompt = [
{
"speaker": "ユーザー",
"text": "Hello, you are an assistant that helps me learn Japanese."
},
{
"speaker": "システム",
"text": "Sure, what can I do for you?"
},
{
"speaker": "ユーザー",
"text": "VRはなんですか。"
}
]
prompt = [
f"{uttr['speaker']}: {uttr['text']}"
for uttr in prompt
]
prompt = "\n".join(prompt)
prompt = (
prompt
+ "\n"
+ "システム: "
)
print(prompt)
"""
ユーザー: Hello, you are an assistant that helps me learn Japanese.
システム: Sure, what can I do for you?
ユーザー: VRはなんですか。
システム:
"""
How to use the model
Notice: Since the model is sensitive to decoding hyper-parameters (e.g. temperature, top_p, top_k, repetition_penalty), it is suggested to explore the best setting for your task.
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("rinna/bilingual-gpt-neox-4b-instruction-ppo", use_fast=False)
model = AutoModelForCausalLM.from_pretrained("rinna/bilingual-gpt-neox-4b-instruction-ppo")
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),
max_new_tokens=512,
do_sample=True,
temperature=1.0,
top_p=0.85,
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):])
print(output)
"""VRとはVirtual Realityの略で、仮想現実とも呼ばれます。これは、コンピューターを使用して仮想世界を作り出し、仮想世界上でコンピューターのゲームや仮想世界を体験するための技術です。この技術は、コンピューターやモバイ ルデバイスの進歩によって、2015年以降、ますます普及しています。VRは、ゲームや仮想世界、その他のアプリケー ションなどのさまざまな分野で、コンピューターと人間の相互作用の新しい方法を提供しています。</s>"""
Tokenization
The model uses a sentencepiece-based tokenizer.
- The tokenizer has a vocabulary size of 65,536.
- It uses byte fallback to decompose unknown text pieces into UTF-8 byte pieces to avoid producing
<UNK>tokens. - It can recognize consecutive whitespaces, newlines, and tabs to handle structured texts better.
- We turned off the default behaviour of prepending leading whitespace because it is not beneficial for processing Japanese.
- Specifically, single whitespace is always processed as one token so that any English word won't have a preceding whitespace like in many other tokenizers (e.g.
_Hello).- This decision trades the English processing efficiency for a unified way to treat whitespaces.
- It leads to a significantly lower loss of next token prediction on English data because whitespaces are easy to predict.
- Don't forget to set
use_fast=Falseto make the above features function correctly.