OpenChat: Less is More for Open-source Models

OpenChat is a series of open-source language models fine-tuned on a diverse and high-quality dataset of multi-round conversations. With only ~6K GPT-4 conversations filtered from the ~90K ShareGPT conversations, OpenChat is designed to achieve high performance with limited data.

Generic models:

• OpenChat: based on LLaMA-13B (2048 context length)
  • 🚀 105.7% of ChatGPT score on Vicuna GPT-4 evaluation
  • 🔥 80.9% Win-rate on AlpacaEval
  • 🤗 Only used 6K data for finetuning!!!
• OpenChat-8192: based on LLaMA-13B (extended to 8192 context length)
  • 106.6% of ChatGPT score on Vicuna GPT-4 evaluation
  • 79.5% Win-rate on AlpacaEval

Code models:

• OpenCoderPlus: based on StarCoderPlus (native 8192 context length)
  • 102.5% of ChatGPT score on Vicuna GPT-4 evaluation
  • 78.7% Win-rate on AlpacaEval

Note: Please load the pretrained models using bfloat16

Code and Inference Server

We provide the full source code, including an inference server compatible with the "ChatCompletions" API, in the OpenChat GitHub repository.

Web UI

OpenChat also includes a web UI for a better user experience. See the GitHub repository for instructions.

Conversation Template

The conversation template involves concatenating tokens.

Besides base model vocabulary, an end-of-turn token <|end_of_turn|> is added, with id eot_token_id.

# OpenChat
[bos_token_id] + tokenize("Human: ") + tokenize(user_question) + [eot_token_id] + tokenize("Assistant: ")
# OpenCoder
tokenize("User:") + tokenize(user_question) + [eot_token_id] + tokenize("Assistant:")

Hint: In BPE, tokenize(A) + tokenize(B) does not always equals to tokenize(A + B)

Following is the code for generating the conversation templates:

@dataclass
class ModelConfig:
    # Prompt
    system: Optional[str]

    role_prefix: dict
    ai_role: str
    eot_token: str
    bos_token: Optional[str] = None

    # Get template
    def generate_conversation_template(self, tokenize_fn, tokenize_special_fn, message_list):
        tokens = []
        masks = []

        # begin of sentence (bos)
        if self.bos_token:
            t = tokenize_special_fn(self.bos_token)
            tokens.append(t)
            masks.append(False)

        # System
        if self.system:
            t = tokenize_fn(self.system) + [tokenize_special_fn(self.eot_token)]
            tokens.extend(t)
            masks.extend([False] * len(t))

        # Messages
        for idx, message in enumerate(message_list):
            # Prefix
            t = tokenize_fn(self.role_prefix[message["from"]])
            tokens.extend(t)
            masks.extend([False] * len(t))

            # Message
            if "value" in message:
                t = tokenize_fn(message["value"]) + [tokenize_special_fn(self.eot_token)]
                tokens.extend(t)
                masks.extend([message["from"] == self.ai_role] * len(t))
            else:
                assert idx == len(message_list) - 1, "Empty message for completion must be on the last."

        return tokens, masks


MODEL_CONFIG_MAP = {
    # OpenChat / OpenChat-8192
    "openchat": ModelConfig(
        # Prompt
        system=None,

        role_prefix={
            "human": "Human: ",
            "gpt": "Assistant: "
        },
        ai_role="gpt",
        eot_token="<|end_of_turn|>",
        bos_token="<s>",
    ),

    # OpenCoder / OpenCoderPlus
    "opencoder": ModelConfig(
        # Prompt
        system=None,

        role_prefix={
            "human": "User:",
            "gpt": "Assistant:"
        },
        ai_role="gpt",
        eot_token="<|end_of_turn|>",
        bos_token=None,
    )
}

License

Our weight license is subject to their corresponding base model. For example, OpenChat and OpenChat-8192 are the same as the model License of LLaMA for non-commercial use only, while OpenCoderPlus is under the License of StarCoder. Furthermore, we should follow the Privacy Practices of ShareGPT. The code released on GitHub is under Apache License 2.0.

Citation

@software{openllms23,
  title = {{OpenLLMs: Less is More for Open-source Models}},
  author = {Wang, Guan and Cheng, Sijie and Yu, Qiying and Liu, Changling},
  doi = {10.5281/zenodo.8105775},
  url = {https://github.com/imoneoi/openchat},
  version = {pre-release},
  year = {2023},
  month = {7},
}