README.md

metadata

language:
  - en
tags:
  - llama

OpenChat: Less is More for Open-source Models

OpenChat is a series of open-source language models fine-tuned on very little diverse and high-quality multi-round conversations. The dataset contains only ~6K GPT-4 conversations filtered from the 90K ShareGPT conversations.

Generic models:

• OpenChat: based on LLaMA-13B (2048 context length)
  • 105.7% of ChatGPT score on Vicuna GPT-4 evaluation
  • 80.87% 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

Code models:

• OpenCoderPlus: based on StarCoderPlus (native 8192 context length)
  • 102.5% of ChatGPT score on Vicuna GPT-4 evaluation
  • 78.70% Win-rate on AlpacaEval

NOTE: Please load the pretrained models using bfloat16

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,
    )
}