presets.py

SYSTEM_PROMPT = "You are a helpful assistant that provide concise and accurate answers."

def set_cora_preset():
    return (
        "gsarti/cora_mgen",  # model_name_or_path
        "<Q>:{current} <P>:{context}",  # input_template
        "<Q>:{current}",  # input_current_text_template
    )


def set_default_preset():
    return (
        "gpt2",  # model_name_or_path
        "{current} {context}",  # input_template
        "{current}",  # output_template
        "{current}",  # contextless_input_template
        "{current}",  # contextless_output_template
        [],  # special_tokens_to_keep
        "",  # decoder_input_output_separator
        "{}",  # model_kwargs
        "{}",  # tokenizer_kwargs
        "{}",  # generation_kwargs
        "{}",  # attribution_kwargs
    )


def set_zephyr_preset():
    return (
        "stabilityai/stablelm-2-zephyr-1_6b",  # model_name_or_path
        "<|system|>{system_prompt}<|endoftext|>\n<|user|>\n{context}\n\n{current}<|endoftext|>\n<|assistant|>".format(system_prompt=SYSTEM_PROMPT),  # input_template
        "<|system|>{system_prompt}<|endoftext|>\n<|user|>\n{current}<|endoftext|>\n<|assistant|>".format(system_prompt=SYSTEM_PROMPT),  # input_current_text_template
        "\n",  # decoder_input_output_separator
        ["<|im_start|>", "<|im_end|>", "<|endoftext|>"],  # special_tokens_to_keep
    )


def set_chatml_preset():
    return (
        "Qwen/Qwen1.5-0.5B-Chat",  # model_name_or_path
        "<|im_start|>system\n{system_prompt}<|im_end|>\n<|im_start|>user\n{context}\n\n{current}<|im_end|>\n<|im_start|>assistant".format(system_prompt=SYSTEM_PROMPT),  # input_template
        "<|im_start|>system\n{system_prompt}<|im_end|>\n<|im_start|>user\n{current}<|im_end|>\n<|im_start|>assistant".format(system_prompt=SYSTEM_PROMPT),  # input_current_text_template
        "\n",  # decoder_input_output_separator
        ["<|im_start|>", "<|im_end|>"],  # special_tokens_to_keep
    )


def set_mmt_preset():
    return (
        "facebook/mbart-large-50-one-to-many-mmt",  # model_name_or_path
        "{context} {current}",  # input_template
        "{context} {current}",  # output_template
        '{\n\t"src_lang": "en_XX",\n\t"tgt_lang": "fr_XX"\n}',  # tokenizer_kwargs
    )


def set_towerinstruct_preset():
    return (
        "Unbabel/TowerInstruct-7B-v0.1",  # model_name_or_path
        "<|im_start|>user\nSource: {current}\nContext: {context}\nTranslate the above text into French. Use the context to guide your answer.\nTarget:<|im_end|>\n<|im_start|>assistant",  # input_template
        "<|im_start|>user\nSource: {current}\nTranslate the above text into French.\nTarget:<|im_end|>\n<|im_start|>assistant",  # input_current_text_template
        "\n",  # decoder_input_output_separator
        ["<|im_start|>", "<|im_end|>"],  # special_tokens_to_keep
    )

def set_gemma_preset():
    return (
        "google/gemma-2b-it", # model_name_or_path
        "<start_of_turn>user\n{context}\n{current}<end_of_turn>\n<start_of_turn>model", # input_template
        "<start_of_turn>user\n{current}<end_of_turn>\n<start_of_turn>model", # input_current_text_template
        "\n", # decoder_input_output_separator
        ["<start_of_turn>", "<end_of_turn>"], # special_tokens_to_keep
    )

def set_mistral_instruct_preset():
    return (
        "mistralai/Mistral-7B-Instruct-v0.2" # model_name_or_path
        "[INST]{context}\n{current}[/INST]" # input_template
        "[INST]{current}[/INST]" # input_current_text_template
        "\n" # decoder_input_output_separator
    )