OpenChat: Advancing Open-source Language Models with Mixed-Quality Data

Online Demo • Discord • Huggingface • Paper

🔥 The first 7B model that Achieves Comparable Results with ChatGPT (March)! 🔥

🤖 #1 Open-source model on MT-bench scoring 7.81, outperforming 70B models 🤖

OpenChat is an innovative library of open-source language models, fine-tuned with C-RLFT - a strategy inspired by offline reinforcement learning. Our models learn from mixed-quality data without preference labels, delivering exceptional performance on par with ChatGPT, even with a 7B model. Despite our simple approach, we are committed to developing a high-performance, commercially viable, open-source large language model, and we continue to make significant strides toward this vision.

News

[2023/11/01] We released the OpenChat-3.5-7B model, surpassing ChatGPT on various benchmarks 🔥.
[2023/09/21] We released our paper OpenChat: Advancing Open-source Language Models with Mixed-Quality Data.
[2023/09/03] We released the OpenChat V3.2 SUPER model.
[2023/08/04] We have launched an Online Demo featuring the latest version, OpenChat 3.2.
[2023/07/30] We are thrilled to introduce the OpenChat V3 model series, based on Llama 2, and now available for free for commercial use!
[2023/07/07] We released the OpenChat V2 model series.
[2023/07/01] We released the OpenChat V1 model series.

Models

Our latest model, OpenChat 3.5, is a highly capable model fine-tuned using C-RLFT with Mistral 7B as the base, on a collection of publicly available high-quality instruction data. For older version models such as OpenChat 3.2 SUPER, please refer to Legacy Models.

To use this model, we highly recommend installing the OpenChat package by following the installation guide and using the OpenChat OpenAI-compatible API server by running the serving command from the table below. The server is optimized for high-throughput deployment using vLLM and can run on a consumer GPU with 24GB RAM. To enable tensor parallelism, append --tensor-parallel-size N to the serving command.

Once started, the server listens at localhost:18888 for requests and is compatible with the OpenAI ChatCompletion API specifications. Please refer to the example request below for reference. Additionally, you can use the OpenChat Web UI for a user-friendly experience.

If you want to deploy the server as an online service, you can use --api-keys sk-KEY1 sk-KEY2 ... to specify allowed API keys and --disable-log-requests --disable-log-stats --log-file openchat.log for logging only to a file. For security purposes, we recommend using an HTTPS gateway in front of the server.

Example request (click to expand)

curl http://localhost:18888/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openchat_3.5",
    "messages": [{"role": "user", "content": "You are a large language model named OpenChat. Write a poem to describe yourself"}]
  }'

Coding Mode

curl http://localhost:18888/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openchat_3.5",
    "condition": "Code",
    "messages": [{"role": "user", "content": "Write an aesthetic TODO app using HTML5 and JS, in a single file. You should use round corners and gradients to make it more aesthetic."}]
  }'

Model	Size	Context	Weights	Serving
OpenChat 3.5	7B	8192	Huggingface	`python -m ochat.serving.openai_api_server --model openchat/openchat_3.5 --engine-use-ray --worker-use-ray`

For inference with Huggingface Transformers (slow and not recommended), follow the conversation template provided below.

Conversation templates (click to expand)

import transformers
tokenizer = transformers.AutoTokenizer.from_pretrained("openchat/openchat_3.5")

# Single-turn
tokens = tokenizer("GPT4 Correct User: Hello<|end_of_turn|>GPT4 Correct Assistant:").input_ids
assert tokens == [1, 420, 6316, 28781, 3198, 3123, 1247, 28747, 22557, 32000, 420, 6316, 28781, 3198, 3123, 21631, 28747]

# Multi-turn
tokens = tokenizer("GPT4 Correct User: Hello<|end_of_turn|>GPT4 Correct Assistant: Hi<|end_of_turn|>GPT4 Correct User: How are you today?<|end_of_turn|>GPT4 Correct Assistant:").input_ids
assert tokens == [1, 420, 6316, 28781, 3198, 3123, 1247, 28747, 22557, 32000, 420, 6316, 28781, 3198, 3123, 21631, 28747, 15359, 32000, 420, 6316, 28781, 3198, 3123, 1247, 28747, 1602, 460, 368, 3154, 28804, 32000, 420, 6316, 28781, 3198, 3123, 21631, 28747]

# Coding Mode
tokens = tokenizer("Code User: Implement quicksort using C++<|end_of_turn|>Code Assistant:").input_ids
assert tokens == [1, 7596, 1247, 28747, 26256, 2936, 7653, 1413, 334, 1680, 32000, 7596, 21631, 28747]

Benchmarks

Model	# Params	Average	MT-Bench	AGIEval	BBH MC	TruthfulQA	MMLU	HumanEval	BBH CoT	GSM8K
OpenChat-3.5	7B	61.6	7.81	47.4	47.6	59.1	64.3	55.5	63.5	77.3
ChatGPT (March)*	?	61.5	7.94	47.1	47.6	57.7	67.3	48.1	70.1	74.9
Mistral	7B	-	6.84	38.0	39.0	-	60.1	30.5	-	52.2
Open-source SOTA**	13B-70B	61.4	7.71	41.7	49.7	62.3	63.7	73.2	41.4	82.3
			WizardLM 70B	Orca 13B	Orca 13B	Platypus2 70B	WizardLM 70B	WizardCoder 34B	Flan-T5 11B	MetaMath 70B

*: ChatGPT (March) results are from GPT-4 Technical Report, Chain-of-Thought Hub, and our evaluation.

**: Open-source SOTA results are taken from reported results in instruction-tuned model papers and official repositories.

***: All zero-shot benchmarks follow the same setting as in the AGIEval paper and Orca paper. CoT tasks use the same configuration as Chain-of-Thought Hub, HumanEval is evaluated with EvalPlus, and MT-bench is run using FastChat. To reproduce our results, follow the instructions below.

Reproducing benchmark results (click to expand)

Reasoning:

Note: Please run the following commands at the base directory of this repository.

python -m ochat.evaluation.run_eval --condition "GPT4 Correct" --model openchat/openchat_3.5
python ochat/evaluation/view_results.py

HumanEval:

Note: Please run the following commands at the base directory of this repository.

python -m ochat.evaluation.run_eval --condition "Code" --eval_sets coding --model openchat/openchat_3.5
python ochat/evaluation/convert_to_evalplus.py

Then all humaneval code samples are placed in ochat/evaluation/evalplus_codegen. Use the following command to evaluate an individual code sample named samples.jsonl using Docker as a sandbox.

docker run -v $(pwd):/app ganler/evalplus:latest --dataset humaneval --samples samples.jsonl

MT-Bench:

Please first launch a local API server, then download FastChat and run the following commands.

Note: Due to non-zero temperature and GPT-4 API changes over time, there might be variations in the results.

cd fastchat/llm_judge
python gen_api_answer.py --model openchat_3.5 --max-tokens 4096 --parallel 128 --openai-api-base http://localhost:18888/v1
python gen_judgment.py --model-list openchat_3.5 --parallel 8 --mode single

Installation

To use OpenChat, you need to install PyTorch, then you can install OpenChat via pip:

pip3 install ochat

If you encounter compatibility problems, you can try to create a new conda environment following the instructions below.

conda create -y --name openchat
conda activate openchat

conda install -y python=3.11
pip3 install torch torchvision torchaudio

pip3 install ochat

In addition to PyPI, you can also install from source (click to expand)

git clone https://github.com/imoneoi/openchat
cd openchat

pip3 install --upgrade pip  # enable PEP 660 support
pip3 install -e .

Web UI

After launching the API server, you can interact with it using OpenChat-UI, which is a fork of Chatbot UI with support for OpenChat models.

To use OpenChat-UI, follow these steps:

Clone the OpenChat-UI repo:

git clone https://github.com/imoneoi/openchat-ui.git

Install Dependencies

npm i

Set the API host to the local server (or the address of the OpenChat server)

Create a .env.local file in the root of the OpenChat-UI repo with the following content:

OPENAI_API_HOST=http://localhost:18888
OPENAI_API_KEY=openchat-dummy-key
NEXT_PUBLIC_DEFAULT_TEMPERATURE=0.5

Run the App

npm run dev

OpenChat Model Training

The OpenChat training system utilizes padding-free training and the Multipack Sampler, achieving a 3~10x speedup compared to the conventional padded training.

Choose a base model

OpenChat supports Llama 2 and Mistral models. Please first choose a base model to fit your needs. Each base model has a corresponding weight repo, model type, and recommended batch size as listed below, they should be filled into BASE_REPO, MODEL_TYPE, and BATCH_SIZE in the following instructions.

Base Model	Size	Weights (with EOT token)	Model Type	Recommended Batch Size per GPU (8xA100 80GB)
Mistral	7B	`imone/Mistral_7B_with_EOT_token`	`openchat_v3.2_mistral`	83968
Llama 2	7B	`imone/LLaMA2_7B_with_EOT_token`	`openchat_v3.2`	83968
Llama 2	13B	`imone/Llama2_13B_with_EOT_token`	`openchat_v3.2`	36864

Note: The OpenChat conversation template requires an <|end_of_turn|> special token. The base model specified must include this token. Our provided weights are the original base weights with this token added. If you want to add them manually, use the convert_llama_weights_to_hf_add_tokens.py or mistral_add_tokens.py in the scripts directory.

Installing DeepSpeed

First, ensure that the CUDA nvcc compiler is available in your environment. If it is not, install the CUDA toolkit that matches the version used by PyTorch.

Next, install DeepSpeed:

pip install deepspeed

Preparing Your Data

To utilize the OpenChat trainer, prepare your SFT data into a JSON Lines format where each line corresponds to a Conversation object:

class Message(BaseModel):
    role: str     # Must be "user" or "assistant"
    content: str  # Message content
    weight: Optional[float] = None  # Loss weight for this message. Typically 0 for user and 1 for assistant to supervise assistant's responses only


class Conversation(BaseModel):
    items: List[Message]  # All messages within the conversation
    condition: str = ""  # C-RLFT condition, can be any string or empty.
    system: str = ""  # System message for this conversation

For basic SFT, assign weight as 0 for human messages and 1 for assistant responses.

SFT example:

{"items":[{"from":"user","content":"Hello","weight":0.0},{"from":"assistant","content":"Hi","weight":1.0},{"from":"user","content":"How are you today?","weight":0.0},{"from":"assistant","content":"I'm fine.","weight":1.0}],"system":""}
{"items":[{"from":"user","content":"Who are you?","weight":0.0},{"from":"assistant","content":"I'm OpenChat.","weight":1.0}],"system":"You are a helpful assistant named OpenChat."}

For C-RLFT, condition should be set as the class the conversation belongs to (e.g. GPT3 or GPT4). The weight is assigned as 0 for human messages and w for assistant responses, where w is the weight of the class (e.g. 0.1 for GPT3 and 1 for GPT4, as found in our C-RLFT paper).

C-RLFT example:

{"items":[{"from":"user","content":"What is C-RLFT?","weight":0.0},{"from":"assistant","content":"C-RLFT is a method for improving open-source LLMs with mixed-quality data.","weight":1.0}],"condition":"GPT4","system":""}
{"items":[{"from":"user","content":"What is C-RLFT?","weight":0.0},{"from":"assistant","content":"I don't know.","weight":0.1}],"condition":"GPT3","system":""}

Pre-tokenizing the Dataset

You'll then need to pre-tokenize the dataset using the command (please specify a filename as PRETOKENIZED_DATA_OUTPUT_PATH to store the pretokenized dataset):

python -m ochat.data.generate_dataset --model-type MODEL_TYPE --model-path BASE_REPO --in-files data.jsonl --out-prefix PRETOKENIZED_DATA_OUTPUT_PATH

Launching the OpenChat Trainer

You can now launch the OpenChat trainer using the command below. Training a 13B model requires eight A/H100s with 80GB VRAM, while a 7B model can be trained with four A/H100s with 80GB VRAM or eight A/H100s with 40GB VRAM.

For hyperparameters, we recommend first setting the batch size to the recommended batch size. If OOM occurs, try setting it to the exact maximum that VRAM can hold and as a multiple of 2048. Other hyperparameters have been carefully selected as the default. Furthermore, the learning rate is automatically determined based on the inverse square-root rule.

Training Commands (click to expand)

NUM_GPUS=8

deepspeed --num_gpus=$NUM_GPUS --module ochat.training_deepspeed.train \
          --model_path BASE_REPO \
          --data_prefix PRETOKENIZED_DATA_OUTPUT_PATH \
          --save_path PATH_TO_SAVE_MODEL \
          --batch_max_len BATCH_SIZE \
          --epochs 5 \
          --save_every 1 \
          --deepspeed \
          --deepspeed_config ochat/training_deepspeed/deepspeed_config.json

You can find checkpoints of all epochs in PATH_TO_SAVE_MODEL. Then you may evaluate each epoch and choose the best one.

Limitations

Foundation Model Limitations Despite its advanced capabilities, OpenChat is still bound by the limitations inherent in its foundation models. These limitations may impact the model's performance in areas such as:

Complex reasoning
Mathematical and arithmetic tasks
Programming and coding challenges

Hallucination of Non-existent Information OpenChat may sometimes generate information that does not exist or is not accurate, also known as "hallucination". Users should be aware of this possibility and verify any critical information obtained from the model.

Safety OpenChat may sometimes generate harmful, hate speech, biased responses, or answer unsafe questions. It's crucial to apply additional AI safety measures in use cases that require safe and moderated responses.

License

Our OpenChat 3.5 code and models are distributed under the Apache License 2.0.

Legacy Models

The following models are older versions of OpenChat and have inferior performance compared to the latest version. They will be deprecated in the next release. Please note that OpenChat V1 and V2 series are now deprecated, please install 3.1.x for using V1 and V2 models

To run the models on multiple GPUs with smaller VRAM, you can enable tensor parallelization, for example, using the --tensor-parallel-size 2 flag.

OpenChat V3 (click to expand)

Model	Size	Context	Weights	Serving
OpenChat 3.2 SUPER	13B	4096	Huggingface	`python -m ochat.serving.openai_api_server --model openchat/openchat_v3.2_super --engine-use-ray --worker-use-ray`

Contact

💌 We are a student team from Tsinghua University, working on OpenChat, a project that requires additional computing power or LLMs API keys for further development. If you are interested in our project and would like to offer support, please feel free to reach out to us:

Wang Guan [imonenext at gmail dot com]
Cheng Sijie [csj23 at mails dot tsinghua dot edu dot cn]

We look forward to hearing from you and collaborating on this exciting project!

Citation

@article{wang2023openchat,
  title={OpenChat: Advancing Open-source Language Models with Mixed-Quality Data},
  author={Wang, Guan and Cheng, Sijie and Zhan, Xianyuan and Li, Xiangang and Song, Sen and Liu, Yang},
  journal={arXiv preprint arXiv:2309.11235},
  year={2023}
}

Acknowledgements

We extend our heartfelt gratitude to Alignment Lab AI, Nous Research, and Pygmalion AI for their substantial contributions to data collection and model training.

Special thanks go to Changling Liu from GPT Desk Pte. Ltd., Qiying Yu at Tsinghua University, Baochang Ma, and Hao Wan from 01.AI company for their generous provision of resources. We are also deeply grateful to Jianxiong Li and Peng Li at Tsinghua University for their insightful discussions.

Furthermore, we appreciate the developers behind the following projects for their significant contributions to our research: Mistral, Chain-of-Thought Hub, Llama 2, Self-Instruct, FastChat (Vicuna), Alpaca, and StarCoder. Their work has been instrumental in driving our research forward.

ftgreat / openchat