DOCSEARCHbyAlgolia
Submit Docs

MediaTek Research Foundation Models

repository·main·Indexed 43 days ago

https://github.com/mtkresearch/mr-models

A collection of foundation models optimized for Traditional Chinese, including the Breeze (Mistral-7B based), BreeXe (Mixtral-8x7B based), and Breeze 2 (LLaMA 3.2 based) families. Features include multi-modal language models, BreezyVoice TTS, and the TC-Eval evaluation suite for Traditional Chinese models.

Tokens
1.3K
Snippets
2
Records
8
Agent score
5%

What's inside MediaTek Research Models

  1. Overview of MediaTek Research Foundation Models

    main
    MediaTek Research (聯發創新基地) develops foundation models specifically optimized for Traditional Chinese users. The research focus includes enhancing linguistic and cultural representation, speech recognition (ASR), speech synthesis (TTS), and multi-modal capabilities (vision-language models).

  2. Overview of BLOOM-zh models (3B and 1B1)

    main

    BLOOM-zh is a collaborative project between Academia Sinica CKIP Lab, MediaTek Research, and the National Academy for Educational Research.

    Usage Restriction: These models are released for non-commercial research purposes only.

    Available Versions:

    • BLOOM-zh-3B: Released on 2023-04-10. It is no longer being updated. Hugging Face Link.
    • BLOOM-zh-1B1: Released on 2023-03-07. It is no longer being updated. Hugging Face Link.

    Technical details and evaluations can be found in this paper.

  4. Explore the Breeze language model family

    main

    Breeze-7B is a language model family built on top of Mistral-7B, optimized for Traditional Chinese.

    • Breeze-7B-Base: Expands the original vocabulary with 30,000 additional Traditional Chinese tokens. This allows it to operate at twice the inference speed for Traditional Chinese compared to Mistral-7B and Llama 7B.
    • Breeze-7B-Instruct: Fine-tuned for tasks such as Q&A, RAG, multi-round chat, and summarization. It shows significant advantages in Traditional Chinese and English benchmarks compared to contemporaries like Taiwan-LLM-7B/13B-chat, QWen(1.5)-7B-Chat, and Yi-6B-Chat.

    Models are available in the Breeze Collection on Hugging Face.

  5. Explore the BreeXe language model family

    main

    Breexe-8x7B is a language model family built on top of Mixtral-8x7B, optimized for Traditional Chinese.

    • Breexe-8x7B-Base: The foundation model. It expands the original vocabulary with 30,000 additional Traditional Chinese tokens, resulting in inference speeds for Traditional Chinese that are twice as fast as the original Mixtral-8x7B.
    • Breexe-8x7B-Instruct: A derivative of the base model fine-tuned for common tasks like Q&A, Retrieval-Augmented Generation (RAG), multi-round chat, and summarization. It performs on par with OpenAI's gpt-3.5-turbo-1106 in Traditional Chinese and English benchmarks.

    Models are available in the Breexe Collection on Hugging Face.

  6. Explore the Breeze 2 multi-modal model family

    main

    Breeze 2 is a suite of advanced multi-modal models based on LLaMA 3.2. It includes:

    • Multi-modal Language Models: Available in 3B and 8B parameter configurations. These models feature vision-aware capabilities (via a visual encoder and bridge module) and support function-calling through prompt templates and post-training.
    • Speech Synthesis (TTS): Known as BreezyVoice, it supports Traditional Chinese, English, and Bopomofo input with Chinese speech output. It also supports voice cloning.

    For implementation details, refer to the BreezyVoice GitHub or the Breeze 2 Collection on Hugging Face.

  7. Capabilities and usage examples for Model 7

    main

    Model 7 is a language model developed by MediaTek Research featuring bilingual capabilities (Traditional Chinese and English) and multi-turn dialogue support. The C-Chat version was released for trial on 2023-09-14.

    Key Capabilities demonstrated in examples:

    • Reading Comprehension: Answering questions based on provided text.
    • Structured Output: Generating content in specific formats like JSON or Markdown tables.
    • Content Generation: Writing formal letters, emails, or creative text.
    • Classification: Classifying text into specific categories (e.g., emotions) and returning results as JSON.
    • Open-ended Question Answering: Discussing complex topics.
    • Role-playing: Adopting specific personas and responding within a defined context.
    • Safety/Ethics: Refusing to provide information on illegal or harmful activities (e.g., manufacturing illegal drugs).
    <<範例：輸出指定格式>>

👩🏻 USER: 生成一本假的中文書籍，並附上它們的相關資訊。以JSON格式表示，需包含以下Key：標題、作者、類型、年份。

💬 MODEL7: {
  "標題": "龍飛鳳舞在翡翠天空",
  "作者": "李明亮",
  "類型": "小說",
  "年份": "2021"
  }

👩🏻 USER: 幫我轉成markdown表格格式

💬 MODEL7:
  |標題            |作者   |類型 | 年份 |
  | ------------- | ----- | --- | --- |
  |龍飛鳳舞在翡翠天空 |李明亮 |小說 | 2021 |