Honyaku-7b-v2

Overview:

Honyaku-7b-v2 is an improved version of its predecessor. This model exhibits enhanced accuracy in adhering to multilingual generation tags compared to the previous version.

Key Features & Limitation:

• Improved Multilingual Generation Accuracy: The model has increased precision in following multilingual generation tags.
• Quality-Reflective Translation: The translation quality of Honyaku-7b is strongly influenced by the pre-training of the base model. Consequently, the quality of translation varies in proportion to the training volume of the original language model.
• The primary purpose is to translate about 500 to several thousand tokens. Due to the characteristics of the Base model, translation into Japanese is the most stable. 　　
• It has been fine-tuned up to 8k tokens, but based on the Base model's characteristics, it supports up to 4k tokens including the prompt.

Cautions:

In minor languages, translation does not function well.
The translation function of 7b-level large language models (LLM) often contains errors.
Do not use unchecked text for social communication.

概要：

Honyaku-7b-v2は、前バージョンの改良版です。このモデルは、多言語生成タグへの追従精度が前バージョンと比較して向上しています。
日本語への翻訳は前バージョンのほうが良い場合があります。

主な特徴と制限事項：

• 多言語生成の精度向上： モデルは、多言語生成タグに対する追従の精度が向上しました。
• 翻訳品質の反映： Honyaku-7bの翻訳品質は、ベースモデルの事前学習に強く影響されます。翻訳品質は、元の言語モデルの学習量に比例して変わります。
• 500～数1000 tokenの翻訳を主目的としています。短すぎる文、長すぎる文で性能低下。
• Base modelの特徴から、日本語への翻訳が最も安定しています。
• 8k tokenまでファインチューニングしていますが、Base modelの特徴からprompt含めて4k tokenにまで対応とします。

注意点：

• マイナーな言語においては、翻訳がうまく機能しません。
• 7bレベルの大規模言語モデル(LLM)の翻訳機能には誤りが多くみられます。未チェックの文章は、正式なコミュニケーションには使用しないでください。

Honyaku-7b-webui

import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, StoppingCriteria, StoppingCriteriaList, TextIteratorStreamer
from threading import Thread

# 言語リスト
languages = [
    "English", "Chinese (Simplified)", "Chinese (Traditional)", "Spanish", "Arabic", "Hindi",
    "Bengali", "Portuguese", "Russian", "Japanese", "German", "French", "Urdu", "Indonesian",
    "Italian", "Turkish", "Korean", "Vietnamese", "Tamil", "Marathi", "Telugu", "Persian",
    "Polish", "Dutch", "Thai", "Gujarati", "Romanian", "Ukrainian", "Malay", "Kannada", "Oriya (Odia)",
    "Burmese (Myanmar)", "Azerbaijani", "Uzbek", "Kurdish (Kurmanji)", "Swedish", "Filipino (Tagalog)",
    "Serbian", "Czech", "Hungarian", "Greek", "Belarusian", "Bulgarian", "Hebrew", "Finnish",
    "Slovak", "Norwegian", "Danish", "Sinhala", "Croatian", "Lithuanian", "Slovenian", "Latvian",
    "Estonian", "Armenian", "Malayalam", "Georgian", "Mongolian", "Afrikaans", "Nepali", "Pashto",
    "Punjabi", "Kurdish", "Kyrgyz", "Somali", "Albanian", "Icelandic", "Basque", "Luxembourgish",
    "Macedonian", "Maltese", "Hawaiian", "Yoruba", "Maori", "Zulu", "Welsh", "Swahili", "Haitian Creole",
    "Lao", "Amharic", "Khmer", "Javanese", "Kazakh", "Malagasy", "Sindhi", "Sundanese", "Tajik", "Xhosa",
    "Yiddish", "Bosnian", "Cebuano", "Chichewa", "Corsican", "Esperanto", "Frisian", "Galician", "Hausa",
    "Hmong", "Igbo", "Irish", "Kinyarwanda", "Latin", "Samoan", "Scots Gaelic", "Sesotho", "Shona",
    "Sotho", "Swedish", "Uyghur"
]

tokenizer = AutoTokenizer.from_pretrained("aixsatoshi/Honyaku-7b-v2")
model = AutoModelForCausalLM.from_pretrained("aixsatoshi/Honyaku-7b-v2", torch_dtype=torch.float16)
model = model.to('cuda:0')

class StopOnTokens(StoppingCriteria):
    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
        stop_ids = [2]
        for stop_id in stop_ids:
            if input_ids[0][-1] == stop_id:
                return True
        return False

def predict(message, history, tokens, temperature, language):
    tag = "<" + language.lower() + ">"
    history_transformer_format = history + [[message, ""]]
    stop = StopOnTokens()

    messages = "".join(["".join(["\n<english>:"+item[0]+"</english>\n", tag+item[1]])
                for item in history_transformer_format])

    model_inputs = tokenizer([messages], return_tensors="pt").to("cuda")
    streamer = TextIteratorStreamer(tokenizer, timeout=10., skip_prompt=True, skip_special_tokens=True)
    generate_kwargs = dict(
        model_inputs,
        streamer=streamer,
        max_new_tokens=int(tokens),
        temperature=float(temperature),
        do_sample=True,
        top_p=0.95,
        top_k=20,
        repetition_penalty=1.15,
        num_beams=1,
        stopping_criteria=StoppingCriteriaList([stop])
    )
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    partial_message = ""
    for new_token in streamer:
        if new_token != '<':
            partial_message += new_token
            yield partial_message

# Gradioインタフェースの設定
demo = gr.ChatInterface(
    fn=predict, 
    title="Honyaku-7b webui",
    description="Translate using Honyaku-7b model",
    additional_inputs=[
        gr.Slider(100, 4096, value=1000, label="Tokens"),
        gr.Slider(0.0, 1.0, value=0.3, label="Temperature"),
        gr.Dropdown(choices=languages, value="Japanese", label="Language")
    ]
)

demo.queue().launch()

Textstreamer

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model_name = "aixsatoshi/Honyaku-7b-v2"
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Define the streamer
streamer = TextStreamer(tokenizer)

# Define the English prompt
english_prompt = """
Machine translation accuracy varies greatly across languages.
Key challenges include context understanding, idiomatic expressions, and syntactic differences.
Advanced models leverage AI to enhance translation quality, focusing on nuances and cultural relevance.

To address these challenges, developers employ neural networks and deep learning techniques, which adapt to linguistic variations and learn from vast amounts of text.
This approach helps in capturing the essence of languages and accurately translating complex sentences.

Furthermore, user feedback plays a crucial role in refining translation algorithms.
By analyzing corrections and suggestions, machine translation systems can evolve and handle nuanced expressions more effectively.
This iterative process ensures continuous improvement, making translations more reliable and understandable for a global audience.
"""

# Prepare the prompt for English to Japanese translation
prompt = f"<english>: {english_prompt} </english>\n\n<japanese>:"

# Tokenize the input text and move to CUDA device
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

# Generate the output using the model and streamer
output = model.generate(**inputs, max_new_tokens=4096, do_sample=True, top_k=20, top_p=0.95, streamer=streamer)

Gradio non-streaming generation

import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

languages = [
    "English", "Chinese (Simplified)", "Chinese (Traditional)", "Spanish", "Arabic", "Hindi",
    "Bengali", "Portuguese", "Russian", "Japanese", "German", "French", "Urdu", "Indonesian",
    "Italian", "Turkish", "Korean", "Vietnamese", "Tamil", "Marathi", "Telugu", "Persian",
    "Polish", "Dutch", "Thai", "Gujarati", "Romanian", "Ukrainian", "Malay", "Kannada", "Oriya (Odia)",
    "Burmese (Myanmar)", "Azerbaijani", "Uzbek", "Kurdish (Kurmanji)", "Swedish", "Filipino (Tagalog)",
    "Serbian", "Czech", "Hungarian", "Greek", "Belarusian", "Bulgarian", "Hebrew", "Finnish",
    "Slovak", "Norwegian", "Danish", "Sinhala", "Croatian", "Lithuanian", "Slovenian", "Latvian",
    "Estonian", "Armenian", "Malayalam", "Georgian", "Mongolian", "Afrikaans", "Nepali", "Pashto",
    "Punjabi", "Kurdish", "Kyrgyz", "Somali", "Albanian", "Icelandic", "Basque", "Luxembourgish",
    "Macedonian", "Maltese", "Hawaiian", "Yoruba", "Maori", "Zulu", "Welsh", "Swahili", "Haitian Creole",
    "Lao", "Amharic", "Khmer", "Javanese", "Kazakh", "Malagasy", "Sindhi", "Sundanese", "Tajik", "Xhosa",
    "Yiddish", "Bosnian", "Cebuano", "Chichewa", "Corsican", "Esperanto", "Frisian", "Galician", "Hausa",
    "Hmong", "Igbo", "Irish", "Kinyarwanda", "Latin", "Samoan", "Scots Gaelic", "Sesotho", "Shona",
    "Sotho", "Swedish", "Uyghur"
]

tokenizer = AutoTokenizer.from_pretrained("aixsatoshi/Honyaku-7b-v2")
model = AutoModelForCausalLM.from_pretrained("aixsatoshi/Honyaku-7b-v2", torch_dtype=torch.float16)
model = model.to('cuda:0')

def predict(message, tokens, temperature, language):
    tag = "<" + language.lower() + ">"
    messages = "\n<english>:" + message + "</english>\n" + tag

    model_inputs = tokenizer([messages], return_tensors="pt").to("cuda")
    output = model.generate(
        **model_inputs,
        max_new_tokens=int(tokens),
        temperature=float(temperature),
        do_sample=True,
        top_p=0.95,
        top_k=20,
        repetition_penalty=1.15,
        num_beams=1,
        eos_token_id=tokenizer.eos_token_id
    )
    translation = tokenizer.decode(output[0], skip_special_tokens=True)
    return translation

# Gradioインタフェースの設定
inputs = [
    gr.Textbox(label="Message", lines=20),
    gr.Slider(100, 4096, value=1000, label="Tokens"),
    gr.Slider(0.0, 1.0, value=0.3, label="Temperature"),
    gr.Dropdown(choices=languages, value="Japanese", label="Language")
]
output = gr.Textbox(label="Translation", lines=35)

demo = gr.Interface(
    fn=predict,
    inputs=inputs,
    outputs=output,
    title="Honyaku-7b webui",
    description="Translate using Honyaku-7b model",
    live=False,  # 明示的にボタンをクリックして翻訳を実行する
    allow_flagging=False
)

demo.launch()

Base Model

tokyotech-llm/Swallow-MS-7b-v0.1