prepare_sft_data.py

# (Modifications Copyright(C) [2024] Advanced Micro Devices, Inc. All rights reserved)
"""
Script for preparing the SFT data for fine-tuning AMD-OLMo model.
Modifed from https://github.com/allenai/OLMo/blob/main/scripts/prepare_tulu_data.py
"""

import logging
from argparse import ArgumentParser
from functools import partial
from pathlib import Path

import datasets as ds
import numpy as np
from rich.progress import track

from olmo.tokenizer import Tokenizer
from olmo.util import prepare_cli_environment
import random
from tqdm import tqdm

log = logging.getLogger(__name__)


def convert_code_feedback_to_tulu_format(dataset, mix=False):
    log.info("Converting code_feedback ...")
    y_all = []
    for i, sample in enumerate(dataset):
        y = {
            "dataset": "code_feedback",
            "id": "code_feedback_{}".format(i),
            "messages": sample['messages']
        }
        y_all.append(y)

    log.info(f"In total {len(y_all)} samples")
    if mix:
        return y_all
    else:
        new_dataset = ds.Dataset.from_list(y_all)
        return new_dataset


def convert_OpenHermes_to_tulu_format(dataset, mix=False):
    log.info("Converting OpenHermes ...")
    role_map = {"human": "user", "gpt": "assistant", "system": "system"}
    y_all = []
    for i, sample in enumerate(dataset):
        y = {
            "dataset": "OpenHermes",
            "id": "OpenHermes_{}".format(i),
            "messages": [{"role": role_map[mssg["from"]], "content": mssg["value"]} for mssg in sample['conversations']]
        }
        y_all.append(y)
    
    log.info(f"In total {len(y_all)} samples")
    if mix:
        return y_all
    else:
        new_dataset = ds.Dataset.from_list(y_all)
        return new_dataset


def convert_WebInstructSub_to_tulu_format(dataset, mix=False):
    log.info("Converting WebInstructSub ...")
    y_all = []
    for i, sample in tqdm(enumerate(dataset)):
        y = {
            "dataset": "WebInstructSub",
            "id": "WebInstructSub_{}".format(i),
            "messages": [{"role": "user", "content": sample["question"]}, {"role": "assistant", "content": sample["answer"]}]
        }
        y_all.append(y)
    
    log.info(f"In total {len(y_all)} samples")
    if mix:
        return y_all
    else:
        new_dataset = ds.Dataset.from_list(y_all)
        return new_dataset
    
    
def main(opts) -> None:
    tokenizer: Tokenizer
    if Path(opts.tokenizer).is_file():
        tokenizer = Tokenizer.from_file(opts.tokenizer, eos_token_id=opts.eos, pad_token_id=opts.pad)
    else:
        tokenizer = Tokenizer.from_pretrained(opts.tokenizer, eos_token_id=opts.eos, pad_token_id=opts.pad)

    if opts.dataset == "tulu":
        dataset = ds.load_dataset("allenai/tulu-v2-sft-mixture", split="train")
    elif opts.dataset == "2nd-phase":
        datasets = ["code-feedback", "OpenHermes", "WebInstructSub"]
        combined_datasets = []
        for dataset_name in datasets:
            if dataset_name == "code-feedback":
                dataset = ds.load_dataset("m-a-p/Code-Feedback", split="train")
                dataset = convert_code_feedback_to_tulu_format(dataset, mix=True)
            elif dataset_name == "OpenHermes":
                dataset = ds.load_dataset("teknium/OpenHermes-2.5", split="train")
                dataset = convert_OpenHermes_to_tulu_format(dataset, mix=True)
            elif dataset_name == "WebInstructSub":
                dataset = ds.load_dataset("TIGER-Lab/WebInstructSub", split="train")
                dataset = convert_WebInstructSub_to_tulu_format(dataset, mix=True)

            combined_datasets += dataset

        random.seed(42)
        random.shuffle(combined_datasets)
        log.info(f"In total {len(combined_datasets)} samples")
        dataset = ds.Dataset.from_list(combined_datasets)

    log.info("Tokenizing dataset...")
    dataset = dataset.map(
        partial(preprocess, tokenizer=tokenizer, max_seq_len=opts.seq_len),
        batched=False,
        remove_columns=["dataset", "id", "messages"],
        num_proc=opts.num_proc,  # type: ignore
    )

    log.info("Filtering dataset...")
    n = len(dataset)  # type: ignore
    dataset = dataset.filter(filter, batched=False, num_proc=opts.num_proc)  # type: ignore
    log.info(f"Filtered out {n - len(dataset):,d} examples")

    log.info("Counting tokens...")
    total_tokens = 0
    for ex in track(dataset):
        assert len(ex["input_ids"]) == opts.seq_len  # type: ignore
        total_tokens += len(ex["input_ids"])  # type: ignore
    log.info(f"Total tokens: {total_tokens:,d}")

    log.info(f"Saving results to '{opts.output_dir}'...")
    output_dir = Path(opts.output_dir)
    output_dir.mkdir(exist_ok=True, parents=True)

    input_ids_file = np.memmap(
        str(output_dir / "input_ids.npy"), dtype=np.uint16, mode="w+", shape=(total_tokens,)
    )
    label_mask_file = np.memmap(
        str(output_dir / "label_mask.npy"), dtype=np.bool_, mode="w+", shape=(total_tokens,)
    )
    offset = 0
    for ex in track(dataset):
        ex_len = len(ex["input_ids"])  # type: ignore
        input_ids_file[offset : offset + ex_len] = ex["input_ids"]  # type: ignore
        label_mask_file[offset : offset + ex_len] = ex["label_mask"]  # type: ignore
        offset += ex_len
    input_ids_file.flush()
    label_mask_file.flush()

    log.info("Done!")


def filter(example):
    return example["n_labels"] > 0


def preprocess(example, tokenizer: Tokenizer, max_seq_len: int):
    input_ids = [tokenizer.eos_token_id]
    label_mask = [False]

    for msg in example["messages"]:
        role_tokens = tokenizer.encode(f"<|{msg['role']}|>\n", add_special_tokens=False)
        label_mask += [False] * len(role_tokens)
        input_ids += role_tokens

        if msg["role"] == "assistant":
            content_tokens = tokenizer.encode(
                msg["content"].strip() + tokenizer.eos_token + "\n", add_special_tokens=False
            )
            label_mask += [True] * len(content_tokens)
            # mask out the last '\n'
            assert content_tokens[-2] == tokenizer.eos_token_id
            label_mask[-1] = False
        else:
            content_tokens = tokenizer.encode(msg["content"].strip() + "\n", add_special_tokens=False)
            label_mask += [False] * len(content_tokens)
        input_ids += content_tokens

    input_ids = input_ids[:max_seq_len]
    label_mask = label_mask[:max_seq_len]

    if len(input_ids) < max_seq_len:
        pad_len = max_seq_len - len(input_ids)
        input_ids += [tokenizer.pad_token_id] * pad_len
        label_mask += [False] * pad_len

    assert len(input_ids) == len(label_mask)
    n_labels = sum(label_mask)

    return {"input_ids": input_ids, "label_mask": label_mask, "n_labels": n_labels}


def get_parser() -> ArgumentParser:
    parser = ArgumentParser(description="Prepare Math dataset")
    parser.add_argument("--output_dir", type=str, help="""Directory to save the results to.""")
    parser.add_argument(
        "-t",
        "--tokenizer",
        type=str,
        help="""Tokenizer path or identifier.""",
        default=Path(__file__).parent / "tokenizers" / "allenai_eleuther-ai-gpt-neox-20b-pii-special.json",
    )
    parser.add_argument("-ds", "--dataset", type=str, help="""Dataset that we are processing. tulu or 2nd-phase""", default="tulu")
    parser.add_argument("-s", "--seq-len", type=int, help="""Max sequence length.""", default=2048)
    parser.add_argument("--eos", type=int, help="""EOS token ID.""", default=50279)
    parser.add_argument("--pad", type=int, help="""PAD token ID.""", default=1)
    parser.add_argument("-j", "--num-proc", type=int, help="""Number of workers.""", default=8)
    return parser


if __name__ == "__main__":
    prepare_cli_environment()
    opts = get_parser().parse_args()
    main(opts)