In the field of artificial intelligence, large language models are developing at an unprecedented speed, and they have become the new favorite of many enterprises, research institutions and individual developers. Today, OpenBuddy, a research team composed of open source enthusiasts and academic researchers in the industry, officially announced the development of the world's first open and commercial Chinese cross-language large model based on the Falcon architecture - OpenBuddy-Falcon-7B!
OpenBuddy: A research team dedicated to opening large language models
The OpenBuddy team has been committed to providing an open, freely available, publicly downloadable, offline-deployable large language model with real cognitive and epiphany capabilities.
Team members include open source enthusiasts and academic researchers in the industry. They devote themselves to contributing to the development of the field of artificial intelligence.
In addition, OpenBuddy actively cooperates with open source culture enthusiasts in academia and industry to jointly promote the development of large language models. Recently, the OpenBuddy team has cooperated with open source enthusiasts in the AI field to jointly explore the upper limit of the capabilities of large language models. Utilizing the content understanding and epiphany capabilities of the OpenBuddy model, NBCE is introduced: an ultra-long context LLM reasoning algorithm. Combined with the OpenBuddy model, NBCE can achieve 10K Context Windows reasoning, which solves the context bottleneck problem of large language models.
Falcon model: the latest SOTA open source large language model
Since the LLaMA model was opened to researchers by Meta, thanks to its excellent performance in various scenarios, LLaMA has become the research basis and evaluation benchmark for many open source projects. However, LLaMA still has problems such as lack of diversity in the training set, undisclosed training data, and license agreements restricting commercial use, which hinder its use in various application scenarios.
In May of this year, Tii Research Institute released the Falcon model, using the commercially available open source protocol of Apache 2.0. The Falcon model is not only more open in terms of protocols and data sets, but also adopts the latest technologies such as Flash Attention, showing amazing performance and content quality.
Once the Falcon model was released, it topped the Huggingface Open LLM Leaderboard and became the latest SOTA open source large language model.
Limitations of the open source model: weak cross-language capabilities
However, existing SOTA open source language models, including LLaMA and Falcon, have a problem: weak cross-language support. These models are mainly based on Indo-European languages such as English and French, and lack in-depth optimization of cross-language dialogue scenarios.
As a result, they have difficulty understanding non-Indo-European languages such as Chinese, let alone having meaningful conversations.
OpenBuddy: a large model with excellent cross-language capabilities
The OpenBuddy team has extensive experience in multilingual model training. It has repeatedly iterated and tuned LLaMA's 7B, 13B, and 30B models, and developed the OpenBuddy-LLaMA series of open source models. These models have excellent question understanding and answering abilities in Chinese, Japanese, Korean and other languages, and can use the knowledge learned from English papers, codes and other materials to provide professional answers and insights for Chinese questions.
After the release of the model in April, OpenBuddy has generated enthusiastic responses in both the open source community and the academic community. In a short period of time, the internal testing community has reached a usage scale of more than a thousand people, and has obtained 295 Stars on GitHub.
As a cross-language model, the OpenBuddy-LLaMA series has excellent question understanding and answering capabilities in Chinese, Japanese, Korean and other languages. OpenBuddy-LLaMA series models have attracted the attention of many open model enthusiasts around the world. Open-source projects such as llama.cpp, Langport, and FastChat have realized the integration of OpenBuddy-LLaMA series models and have been widely used in the community.
OpenBuddy-Falcon-7B: A cross-language large model based on the Falcon architecture
The OpenBuddy team was very interested in the Falcon model from the very beginning. After unremitting research and trial and error, the OpenBuddy team successfully mastered the training know-how of the Falcon model, and launched the world's first open and commercial Chinese cross-language large model based on the Falcon architecture - OpenBuddy-Falcon- 7B .
OpenBuddy-Falcon-7B has the following advantages in terms of cross-language capabilities:
Multiple languages are supported, including Chinese, Japanese, Korean, English, French, German, etc.
A novel cross-language learning technology is adopted, which greatly improves the model's ability to understand and integrate cross-language knowledge.
In-depth optimization has been carried out for the dialogue scene, allowing the model to conduct free and smooth dialogues between various languages. The model can understand the needs of users and complete various tasks such as question and answer, writing, imitation, and creation. At the same time, the model has excellent Cross-lingual ability, can directly switch between multiple languages, and complete translation and other tasks that require cross-language ability.
Capable of using plug-ins, the model can analyze user intentions, automatically select and call plug-ins. For example, call the Stable Diffusion open source model drawing, call the built-in search engine to search for information, and so on. The model can deeply understand the needs of users, instead of repeating literal requirements, and can complete complex instructions such as "draw a street full of traffic" and "draw an animal that an ugly duckling grows into".
Open, commercially available cross-language large models: a new era
The release of OpenBuddy-Falcon-7B marks the arrival of a new era. In this era, cross-language large models are no longer the patent of technology giants, but open and commercially available resources, which can provide strong support for developers, enterprises and researchers around the world.
It is believed that the release of OpenBuddy-Falcon-7B will have a profound impact on the development of the field of cross-language AI. In the days to come, we are expected to witness the advent of more innovative applications and breakthrough technologies based on OpenBuddy-Falcon-7B, jointly promoting the prosperity and development of the field of artificial intelligence.
Try OpenBuddy-Falcon-7B to start your AI innovation journey
Welcome to the OpenBuddy official website: https://OpenBuddy.ai, experience the OpenBuddy-Falcon-7B model, and freely deploy and commercialize it according to the Apache 2.0 protocol. Let's explore together the endless possibilities brought by the OpenBuddy-Falcon series of models.
For more detailed information about OpenBuddy-Falcon series models, please pay attention to OpenBuddy official website and articles. May you be full of surprises and achievements in your journey of AI innovation!
The future: 40 billion large models
In addition to the model with 7B (7 billion) parameters, OpenBuddy also has its own insights and experience in the training of tens of billions of models. OpenBuddy-LLaMA-30B is still in the internal testing stage, and has achieved test results close to or even slightly better than ChatGPT-3.5 in a double-blind test team composed of volunteers.
Currently, OpenBuddy is training large models with 13B (13 billion) and 40B (40 billion) parameters: OpenBuddy-LLaMA-13B, OpenBuddy-Falcon-40B, which are expected to emerge with stronger cognitive and epiphany abilities. The model is still being perfected, and it is impossible to completely avoid the generation of harmful content in the model. If the developer studies and uses it by himself, he needs to pay attention.
https://github.com/OpenBuddy/OpenBuddy
