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

💻Online Demo | 🤗Huggingface | 📃Paper | 💭Discord

• 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 which can be run on a consumer GPU (e.g. RTX 3090).
• 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

• [2024/05/22] We released the Llama-3 based version OpenChat 3.6 20240522, outperforming official Llama 3 8B Instruct and open-source finetunes/merges.

• [2024/01/06] We released the second update, OpenChat 3.5 0106, further improved coding and overall performance 🏆.

• [2023/12/10] We released the first update, OpenChat 3.5 1210, improved coding by 15 points 🚀.

• [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.

🏷️ Benchmarks - OpenChat 3.6

python -m ochat.evaluation.run_eval --condition "GPT4 Correct" --model openchat/openchat-3.6-8b-20240522 --eval_sets fs_cothub/mmlu fs_cothub/gsm8k fs_cothub/math
python -m ochat.evaluation.run_eval --condition "GPT4" --model openchat/openchat-3.6-8b-20240522 --eval_sets zs/gpqa

🏷️ Benchmarks - OpenChat 3.5

🔥 OpenChat-3.5-0106 (7B) now outperforms Grok-0 (33B) on all 4 benchmarks and Grok-1 (314B) on average and 3/4 benchmarks.

python -m ochat.evaluation.run_eval --condition "GPT4 Correct" --model openchat/openchat-3.5-0106 --eval_sets coding fs_cothub/bbh fs_cothub/mmlu zs/agieval zs/bbh_mc_orca zs/truthfulqa_orca
python ochat/evaluation/view_results.py
python ochat/evaluation/convert_to_evalplus.py

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

python -m ochat.evaluation.run_eval --condition "Math Correct" --model openchat/openchat-3.5-0106 --eval_sets fs_cothub/gsm8k zs/math
python ochat/evaluation/view_results.py

cd fastchat/llm_judge
python gen_api_answer.py --model openchat-3.5-0106 --max-tokens 4096 --parallel 128 --openai-api-base http://localhost:18888/v1
python gen_judgment.py --model-list openchat-3.5-0106 --parallel 8 --mode single

⬇️ Installation

[!NOTE] Need pytorch and CUDA to run OpenChat

pip

pip3 install ochat

[!IMPORTANT] If you are facing package compatibility issues with pip, try the conda method below or check this issue

conda

conda create -y --name openchat python=3.11
conda activate openchat

pip3 install ochat

Windows (WSL 1.x, Ubuntu-22.04)

sudo apt update
sudo apt install build-essential

sudo apt install -y curl
curl -o miniconda.sh https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash miniconda.sh

# Restart WSL terminal if the following conda command does not work

conda create -y --name openchat python=3.11
conda activate openchat

pip3 install ochat

From source

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

pip3 install --upgrade pip  # enable PEP 660 support
pip3 install -e .  # Editable mode, you can make changes in this cloned repo

🚀 Deploying API server

⚡ Our API server is ready for production use and compatible with the OpenAI API protocol. It is highly optimized with vLLM and can dynamically batch requests.

📎 Note: For 20 series or older GPUs that do not support bfloat16, add --dtype float16 to the server args.

List of currently supported models

For a single GPU (e.g. RTX 3090, 4090)

python -m ochat.serving.openai_api_server --model MODEL_REPO

For multiple GPUs (tensor parallel)

# N is the number of tensor parallel GPUs
python -m ochat.serving.openai_api_server --model MODEL_REPO --engine-use-ray --worker-use-ray --tensor-parallel-size N

use -h to see more settings

python -m ochat.serving.openai_api_server --model MODEL_REPO -h

Request example

Once started, the server listens at localhost:18888 for requests and is compatible with the OpenAI ChatCompletion API specifications.

💡 Default Mode (GPT4 Correct): Best for coding, chat and general tasks

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

🧮 Mathematical Reasoning Mode: Tailored for solving math problems

curl http://localhost:18888/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MODEL_TYPE",
    "condition": "Math Correct",
    "messages": [{"role": "user", "content": "10.3 − 7988.8133 = "}]
  }'

🌐 Web UI - OpenChat-UI

After launching the API server, OpenChat provide user interface that easy to interact with. Click here to check Web UI

🤗 Inference with Transformers

[!WARNING] It's recommended to use our optimized API server for deployment. Inferencing with Transformers will be slower.

💡 Default Mode (GPT4 Correct): Best for coding, chat and general tasks

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:

🧮 Mathematical Reasoning Mode: Tailored for solving math problems

Math Correct User: 10.3 − 7988.8133=<|end_of_turn|>Math Correct Assistant:

⚠️ Notice: Remember to set <|end_of_turn|> as end of generation token.

The default (GPT4 Correct) template is also available as the integrated tokenizer.chat_template, which can be used instead of manually specifying the template.

🛠️ 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 3 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.

Note: The OpenChat conversation template requires <|eot_id|>, <|start_header_id|>, <|end_header_id|> (Llama 3) <|end_of_turn|> (Mistral) special tokens. The base model specified must include these tokens with initialized embeddings. Our provided weights are the original base weights with this token added and embeddings initialized. If you want to add them manually, use the init_special_embedding_llama3.py or mistral_add_tokens.py in the scripts directory.

Installing DeepSpeed and Flash Attention

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 building dependencies:

pip install packaging ninja

Finally, install the packages:

pip install deepspeed flash-attn

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":[{"role":"user","content":"Hello","weight":0.0},{"role":"assistant","content":"Hi","weight":1.0},{"role":"user","content":"How are you today?","weight":0.0},{"role":"assistant","content":"I'm fine.","weight":1.0}],"system":""}
{"items":[{"role":"user","content":"Who are you?","weight":0.0},{"role":"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":[{"role":"user","content":"What is C-RLFT?","weight":0.0},{"role":"assistant","content":"C-RLFT is a method for improving open-source LLMs with mixed-quality data.","weight":1.0}],"condition":"GPT4","system":""}
{"items":[{"role":"user","content":"What is C-RLFT?","weight":0.0},{"role":"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.

• 13B model requires eight A/H100s with 80GB VRAM
• 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.

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 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

Code is distributed under the Apache License 2.0.

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}
}

💌Contact

Project Lead:

• Guan Wang [imonenext at gmail dot com]
• Alpay Ariyak [aariyak at wpi dot edu]

Main Contributors:

• Xianyuan Zhan (Tsinghua University)
• Qiying Yu (Tsinghua University)
• Changling Liu (GPT Desk Pte. Ltd.)
• LDJ
• AutoMeta (Alignment Lab AI)

Sponsors:

• Sen Song (Tsinghua University)
• Yang Liu (Tsinghua University)
• 01.AI Company
• RunPod

Special Thanks:

• Mistral
• Chain-of-Thought Hub
• Llama 2
• Self-Instruct
• FastChat (Vicuna)
• Alpaca
• StarCoder