LLM Zoo: democratizing ChatGPT

⚡LLM Zoo is a project that provides data, models, and evaluation benchmark for large language models.⚡ [Tech Report]

✨ Latest News

• [07/12/2023]: More instruction-following data of different languages is available here.
• [05/05/2023]: Release the training code. Now, you can replicate a multilingual instruction-following LLM by yourself. :-)
• [04/24/2023]: Add more results (e.g., MOSS) in the evaluation benchmark.
• [04/08/2023]: Release the Phoenix (for all languages) and Chimera (for Latin languages) models.

🤔 Motivation

• Break "AI supremacy" and democratize ChatGPT

"AI supremacy" is understood as a company's absolute leadership and monopoly position in an AI field, which may even include exclusive capabilities beyond general artificial intelligence. This is unacceptable for AI community and may even lead to individual influence on the direction of the human future, thus bringing various hazards to human society.

• Make ChatGPT-like LLM accessible across countries and languages
• Make AI open again. Every person, regardless of their skin color or place of birth, should have equal access to the technology gifted by the creator. For example, many pioneers have made great efforts to spread the use of light bulbs and vaccines to developing countries. Similarly, ChatGPT, one of the greatest technological advancements in modern history, should also be made available to all.

🎬 Get started

Install

Run the following command to install the required packages:

pip install -r requirements.txt

CLI Inference

python -m llmzoo.deploy.cli --model-path /path/to/weights/

For example, for Phoenix, run

python -m llmzoo.deploy.cli --model-path FreedomIntelligence/phoenix-inst-chat-7b

and it will download the model from Hugging Face automatically. For Chimera, please follow this instruction to prepare the weights.

Check here for deploying a web application.

📚 Data

Overview

We used the following two types of data for training Phoenix and Chimera:

Instruction data

• Multilingual instructions (language-agnostic instructions with post-translation)

+ Self-Instructed / Translated (Instruction, Input) in Language A
- ---(Step 1) Translation --->
+ (Instruction, Input) in Language B (B is randomly sampled w.r.t. the probability distribution of realistic languages)
- ---(Step 2) Generate--->
+ Output in Language B

• User-centered instructions

+ (Role, Instruction, Input) seeds
- ---(Step 1) Self Instruct--->
+ (Role, Instruction, Input) samples
- ---(Step 2) generate output Instruct--->
+ (Role, Instruction, Input) ---> Output

Conversation data

• User-shared conversations

+ ChatGPT conversations shared on the Internet
- ---(Step 1) Crawl--->
+ Multi-round conversation data

Check InstructionZoo for the collection of instruction datasets.

Check GPT-API-Accelerate Tool for faster data generation using ChatGPT.

Download

• phoenix-sft-data-v1: The data used for training Phoenix and Chimera.

🐼 Models

Overview of existing models

Model	Backbone	#Params	Open-source model	Open-source data	Claimed language	Post-training (instruction)	Post-training (conversation)	Release date
ChatGPT	-	-	❌	❌	multi			11/30/22
Wenxin	-	-	❌	❌	zh			03/16/23
ChatGLM	GLM	6B	✅	❌	en, zh			03/16/23
Alpaca	LLaMA	7B	✅	✅	en	52K, en	❌	03/13/23
Dolly	GPT-J	6B	✅	✅	en	52K, en	❌	03/24/23
BELLE	BLOOMZ	7B	✅	✅	zh	1.5M, zh	❌	03/26/23
Guanaco	LLaMA	7B	✅	✅	en, zh, ja, de	534K, multi	❌	03/26/23
Chinese-LLaMA-Alpaca	LLaMA	7/13B	✅	✅	en, zh	2M/3M, en/zh	❌	03/28/23
LuoTuo	LLaMA	7B	✅	✅	zh	52K, zh	❌	03/31/23
Vicuna	LLaMA	7/13B	✅	✅	en	❌	70K, multi	03/13/23
Koala	LLaMA	13B	✅	✅	en	355K, en	117K, en	04/03/23
BAIZE	LLaMA	7/13/30B	✅	✅	en	52K, en	111.5K, en	04/04/23
Phoenix (Ours)	BLOOMZ	7B	✅	✅	multi	40+	40+	04/08/23
Latin Phoenix: Chimera (Ours)	LLaMA	7/13B	✅	✅	multi (Latin)	Latin	Latin	04/08/23

The key difference between existing models and ours.

The key difference in our models is that we utilize two sets of data, namely instructions and conversations, which were previously only used by Alpaca and Vicuna respectively. We believe that incorporating both types of data is essential for a recipe to achieve a proficient language model. The rationale is that the instruction data helps to tame language models to adhere to human instructions and fulfill their information requirements, while the conversation data facilitates the development of conversational skills in the model. Together, these two types of data complement each other to create a more well-rounded language model.

Phoenix (LLM across Languages)

The philosophy to name

The first model is named Phoenix. In Chinese culture, the Phoenix is commonly regarded as a symbol of the king of birds; as the saying goes "百鸟朝凤", indicating its ability to coordinate with all birds, even if they speak different languages. We refer to Phoenix as the one capable of understanding and speaking hundreds of (bird) languages. More importantly, Phoenix is the totem of "the Chinese University of Hong Kong, Shenzhen" (CUHKSZ); it goes without saying this is also for the Chinese University of Hong Kong (CUHK).

Model	Backbone	Data	Link
Phoenix-chat-7b	BLOOMZ-7b1-mt	Conversation	parameters
Phoenix-inst-chat-7b	BLOOMZ-7b1-mt	Instruction + Conversation	parameters
Phoenix-inst-chat-7b-int4	BLOOMZ-7b1-mt	Instruction + Conversation	parameters

Chimera (LLM mainly for Latin and Cyrillic languages)

The philosophy to name

The philosophy to name: The biggest barrier to LLM is that we do not have enough candidate names for LLMs, as LLAMA, Guanaco, Vicuna, and Alpaca have already been used, and there are no more members in the camel family. Therefore, we find a similar hybrid creature in Greek mythology, Chimera, composed of different Lycia and Asia Minor animal parts. Coincidentally, it is a hero/role in DOTA (and also Warcraft III). It could therefore be used to memorize a period of playing games overnight during high school and undergraduate time.

Model	Backbone	Data	Link
Chimera-chat-7b	LLaMA-7b	Conversation	parameters (delta)
Chimera-chat-13b	LLaMA-13b	Conversation	parameters (delta)
Chimera-inst-chat-7b	LLaMA-7b	Instruction + Conversation	parameters (delta)
Chimera-inst-chat-13b	LLaMA-13b	Instruction + Conversation	parameters (delta)

Due to LLaMA's license restrictions, we follow FastChat to release our delta weights. To use Chimera, download the original LLaMA weights and run the script:

python tools/apply_delta.py \
 --base /path/to/llama-13b \
 --target /output/path/to/chimera-inst-chat-13b \
 --delta FreedomIntelligence/chimera-inst-chat-13b-delta

CAMEL (Chinese And Medically Enhanced Langauge models)

The philosophy to name

The philosophy to name: Its Chinese name is HuatuoGPT or 华佗GPT to commemorate the great Chinese physician named Hua Tuo (华佗), who lived around 200 AC. Training is already finished; we will release it in two weeks; some efforts are needed to deploy it in public cloud servers in case of massive requests.

Check our models in HuatuoGPT or try our demo . Similar biomedical models could be seen in biomedical LLMs.

More models in the future

Legal GPT (coming soon)

Vision-Language Models (coming soon)

Retrieval-augmented Models (coming soon)

🧐 Evaluation and Benchmark

We provide a bilingual, multidimensional comparison across different open-source models with ours.

Chinese

• Automatic Evaluation Using GPT-4:

Model	Ratio
Phoenix-inst-chat-7b vs. ChatGPT	85.2\%
Phoenix-inst-chat-7b vs. ChatGLM-6b	94.6\%
Phoenix-inst-chat-7b vs. Baidu-Wenxin	96.8\%
Phoenix-inst-chat-7b vs. MOSS-moon-003-sft	109.7\%
Phoenix-inst-chat-7b vs. BELLE-7b-2m	122.7\%
Phoenix-inst-chat-7b vs. Chinese-Alpaca-7b	135.3\%
Phoenix-inst-chat-7b vs. Chinese-Alpaca-13b	125.2\%

Observation: It shows that Phoenix-chat-7b achieves 85.2\% performance of ChatGPT in Chinese. It slightly underperforms Baidu-Wenxin (96.8\%) and ChatGLM-6b (94.6 \%), both are not fully open-source; ChatGLM-6b only provides model weights without training data and details. Although Phoenix is a multilingual LLM, it achieves SOTA performance among all open-source Chinese LLMs.

• Human Evaluation:

	win	tie	lose
Phoenix vs. ChatGPT	12	35	53
Phoenix vs. ChatGLM-6b	36	11	53
Phoenix vs. Baidu-Wenxin	29	25	46
Phoenix vs. BELLE-7b-2m	55	31	14
Phoenix vs. Chinese-Alpaca-13b	56	31	13

Observation: It shows that the human evaluation results show the same trend as the automatic evaluation results.

English

• Automatic Evaluation Using GPT-4:

Model	Ratio
Chimera-chat-7b vs. ChatGPT	85.2\%
Chimera-chat-13b vs. ChatGPT	92.6\%
Chimera-inst-chat-13b vs. ChatGPT	96.6\%

👾 Quantization

We offer int8 and int4 quantizati