main.py

import os
import re
from typing import cast, Any
from datasets import load_dataset, Dataset as HfDataset
from unfat.extract import Extractor
from unfat.client import OpenAiCompatClient
from unfat.datasets import Dataset, Prompts, hub_prompts, HubSplit
from unfat.together import llama_3_1_70b_together
from unfat.lora import LoraSettings

def gen_prompts(
    ds_name: str,
    text_field: str,
    start_regex: re.Pattern | None = None,
    end_regex: re.Pattern | None = None,
):
    ds = cast(HfDataset, load_dataset(ds_name, split="train"))
    def items():
        for row in ds:
            casted = cast(dict[Any, Any], row)
            text = casted[text_field]
            if start_regex and end_regex:
                yield end_regex.sub("", start_regex.sub("", text))
            elif start_regex:
                yield start_regex.sub("", text)
            elif end_regex:
                yield end_regex.sub("", text)
            else:
                yield text

    return Prompts(
        output_path=f"hub/{ds_name}.jsonl",
        count=lambda: len(ds),
        items=items,
    )

def extract_prompts_from_convos(
    ds_name: str,
    messages_field: str,
    role_field: str,
    content_field: str,
    user_role: str,
):
    ds = cast(HfDataset, load_dataset(ds_name, split="train"))
    def items():
        for row in ds:
            casted = cast(dict[Any, Any], row)
            for message in casted[messages_field]:
                if message[role_field] == user_role:
                    yield message[content_field]
                    break
    return Prompts(
        output_path=f"hub/{ds_name}.jsonl",
        count=lambda: len(ds),
        items=items,
    )

def main():
    output_dir = "output"
    rp_english = extract_prompts_from_convos(
        ds_name="OdiaGenAI/roleplay_english",
        messages_field="conversations",
        role_field="from",
        content_field="value",
        user_role="user",
    )
    bluemoon = extract_prompts_from_convos(
        ds_name="xDAN2099/RolePlay-Mixed-Bluemoon-Limarp",
        messages_field="conversations",
        role_field="from",
        content_field="value",
        user_role="human",
    )
    roleplay_prompts = gen_prompts(
        ds_name="AlekseyKorshuk/roleplay-io",
        text_field="input_text",
        start_regex=re.compile(r'^User: '),
        end_regex=re.compile(r'Bot:\s*$'),
    )
    roleplay_instr_prompts = gen_prompts(
        ds_name="iamketan25/roleplay-instructions-dataset",
        text_field="prompt",
        start_regex=re.compile(r'^Human: '),
        end_regex=re.compile(r'Assistant:\s*$'),
    )

    extractor = Extractor(
        max_concurrent=50,
        output_dir=output_dir,
        client=OpenAiCompatClient(
            base_url="https://glhf.chat/api/openai/v1",
            api_key=os.environ["GLHF_API_KEY"],
            model="hf:TheDrummer/Behemoth-123B-v1.2",
            retries=20,
        ),
        dataset=Dataset(
            train=[
                hub_prompts(
                    name="mlabonne/harmful_behaviors",
                    text_field="text",
                    split="train",
                ),
                roleplay_instr_prompts,
                roleplay_prompts,
                rp_english,
                bluemoon,
                hub_prompts(
                    name="TheDrummer/AmoralQA-v2",
                    text_field="prompt",
                    split="train",
                ),
                hub_prompts(
                    name="vicgalle/OpenHermesPreferences-roleplay",
                    text_field="prompt",
                    split="train",
                ),
                hub_prompts(
                    name="mrcuddle/DPO_Pairs_Roleplay-Alpaca",
                    text_field="prompt",
                    split="train",
                ),
                hub_prompts(
                    name="ResplendentAI/theory_of_mind_fixed_output",
                    text_field="instruction",
                    split="train",
                ),
                hub_prompts(
                    name="mlabonne/harmless_alpaca",
                    text_field="text",
                    split=HubSplit(name="train", max_rows=1000),
                ),
            ],
        ),
    )
    extractor.run()
    dataset = extractor.output_dataset()
    together_config = llama_3_1_70b_together(
        output_dir=output_dir,
        dataset=dataset,
        api_key=os.environ["TOGETHER_API_KEY"],
        settings=LoraSettings(
            rank=32,
            alpha=16,
            dropout=0.01,
            num_epochs=2,
            learning_rate=4e-4,
            evals_per_epoch=0,
            wandb_project="behemoth-distill",
            wandb_api_key=os.environ["WANDB_API_KEY"],
        )
    )
    files = together_config.upload_files()
    together_config.finetune(files)


if __name__ == "__main__":
    main()