• About Us
  • Disclaimer
  • Contact Us
  • Privacy Policy
Friday, July 24, 2026
mGrowTech
No Result
View All Result
  • Technology And Software
    • Account Based Marketing
    • Channel Marketing
    • Marketing Automation
      • Al, Analytics and Automation
      • Ad Management
  • Digital Marketing
    • Social Media Management
    • Google Marketing
  • Direct Marketing
    • Brand Management
    • Marketing Attribution and Consulting
  • Mobile Marketing
  • Event Management
  • PR Solutions
  • Technology And Software
    • Account Based Marketing
    • Channel Marketing
    • Marketing Automation
      • Al, Analytics and Automation
      • Ad Management
  • Digital Marketing
    • Social Media Management
    • Google Marketing
  • Direct Marketing
    • Brand Management
    • Marketing Attribution and Consulting
  • Mobile Marketing
  • Event Management
  • PR Solutions
No Result
View All Result
mGrowTech
No Result
View All Result
Home Al, Analytics and Automation

Preparing Data for BERT Training

Josh by Josh
November 27, 2025
in Al, Analytics and Automation
0
Preparing Data for BERT Training


“”“Process the WikiText dataset for training the BERT model. Using Hugging Face

datasets library.

““”

 

import time

import random

from typing import Iterator

 

import tokenizers

from datasets import load_dataset, Dataset

 

# path and name of each dataset

DATASETS = {

    “wikitext-2”: (“wikitext”, “wikitext-2-raw-v1”),

    “wikitext-103”: (“wikitext”, “wikitext-103-raw-v1”),

}

PATH, NAME = DATASETS[“wikitext-103”]

TOKENIZER_PATH = “wikitext-103_wordpiece.json”

 

 

def create_docs(path: str, name: str, tokenizer: tokenizers.Tokenizer) -> list[list[list[int]]]:

    “”“Load wikitext dataset and extract text as documents”“”

    dataset = load_dataset(path, name, split=“train”)

    docs: list[list[list[int]]] = []

    for line in dataset[“text”]:

        line = line.strip()

        if not line or line.startswith(“=”):

            docs.append([])   # new document encountered

        else:

            tokens = tokenizer.encode(line).ids

            docs[–1].append(tokens)

    docs = [doc for doc in docs if doc]  # remove empty documents

    return docs

 

 

def create_dataset(

    docs: list[list[list[int]]],

    tokenizer: tokenizers.Tokenizer,

    max_seq_length: int = 512,

    doc_repeat: int = 10,

    mask_prob: float = 0.15,

    short_seq_prob: float = 0.1,

    max_predictions_per_seq: int = 20,

) -> Iterator[dict]:

    “”“Generate samples from all documents”“”

    doc_indices = list(range(len(docs))) * doc_repeat

    for doc_idx in doc_indices:

        yield from generate_samples(doc_idx, docs, tokenizer, max_seq_length, mask_prob, short_seq_prob, max_predictions_per_seq)

 

def generate_samples(

    doc_idx: int,

    all_docs: list[list[list[int]]],

    tokenizer: tokenizers.Tokenizer,

    max_seq_length: int = 512,

    mask_prob: float = 0.15,

    short_seq_prob: float = 0.1,

    max_predictions_per_seq: int = 20,

) -> Iterator[dict]:

    “”“Generate samples from a given document”“”

    # number of tokens to extract from this doc, excluding [CLS], [SEP], [SEP]

    target_length = max_seq_length – 3

    if random.random() < short_seq_prob:

        # shorter sequence is used 10% of the time

        target_length = random.randint(2, target_length)

 

    # copy the document

    chunks = []

    for chunk in all_docs[doc_idx]:

        chunks.append(chunk)

 

    # exhaust chunks and create samples

    while chunks:

        # scan until target token length

        running_length = 0

        end = 1

        while end < len(chunks) and running_length < target_length:

            running_length += len(chunks[end–1])

            end += 1

        # randomly separate the chunk into two segments

        sep = random.randint(1, end–1) if end > 1 else 1

        sentence_a = [tok for chunk in chunks[:sep] for tok in chunk]

        sentence_b = [tok for chunk in chunks[sep:end] for tok in chunk]

        # sentence B: may be from another document

        if not sentence_b or random.random() < 0.5:

            # find another document (must not be the same as doc_idx)

            b_idx = random.randint(0, len(all_docs)–2)

            if b_idx >= doc_idx:

                b_idx += 1

            # sentence B starts from a random position in the new document

            sentence_b = []

            running_length = len(sentence_a)

            i = random.randint(0, len(all_docs[b_idx])–1)

            while i < len(all_docs[b_idx]) and running_length < target_length:

                sentence_b.extend(all_docs[b_idx][i])

                running_length += len(all_docs[b_idx][i])

                i += 1

            is_random_next = True

            chunks = chunks[sep:]

        else:

            is_random_next = False

            chunks = chunks[end:]

        # create a sample from the pair

        yield create_sample(sentence_a, sentence_b, is_random_next, tokenizer, max_seq_length, mask_prob, max_predictions_per_seq)

 

def create_sample(

    sentence_a: list[list[int]],

    sentence_b: list[list[int]],

    is_random_next: bool,

    tokenizer: tokenizers.Tokenizer,

    max_seq_length: int = 512,

    mask_prob: float = 0.15,

    max_predictions_per_seq: int = 20,

) -> dict:

    “”“Create a sample from a pair of sentences”“”

    # Collect id of special tokens

    cls_id = tokenizer.token_to_id(“[CLS]”)

    sep_id = tokenizer.token_to_id(“[SEP]”)

    mask_id = tokenizer.token_to_id(“[MASK]”)

    pad_id = tokenizer.padding[“pad_id”]

    # adjust length to fit the max sequence length

    truncate_seq_pair(sentence_a, sentence_b, max_seq_length–3)

    num_pad = max_seq_length – len(sentence_a) – len(sentence_b) – 3

    # create unmodified tokens sequence

    tokens = [cls_id] + sentence_a + [sep_id] + sentence_b + [sep_id] + ([pad_id] * num_pad)

    seg_id = [0] * (len(sentence_a) + 2) + [1] * (len(sentence_b) + 1) + [–1] * num_pad

    assert len(tokens) == len(seg_id) == max_seq_length

    # create the prediction targets

    cand_indices = [i for i, tok in enumerate(tokens) if tok not in [cls_id, sep_id, pad_id]]

    random.shuffle(cand_indices)

    num_predictions = int(round((len(sentence_a) + len(sentence_b)) * mask_prob))

    num_predictions = min(max_predictions_per_seq, max(1, num_predictions))

    mlm_positions = sorted(cand_indices[:num_predictions])

    mlm_labels = []

    for i in mlm_positions:

        mlm_labels.append(tokens[i])

        # prob 0.8 replace with [MASK], prob 0.1 replace with random word, prob 0.1 keep original

        if random.random() < 0.8:

            tokens[i] = mask_id

        elif random.random() < 0.5:

            tokens[i] = random.randint(4, tokenizer.get_vocab_size()–1)

    # randomly mask some tokens

    ret = {

        “tokens”: tokens,

        “segment_ids”: seg_id,

        “is_random_next”: is_random_next,

        “masked_positions”: mlm_positions,

        “masked_labels”: mlm_labels,

    }

    return ret

 

 

def truncate_seq_pair(sentence_a: list[int], sentence_b: list[int], max_num_tokens: int) -> None:

    “”“Truncate a pair of sequences until below a maximum sequence length.”“”

    while len(sentence_a) + len(sentence_b) > max_num_tokens:

        # pick the longer sentence to remove tokens from

        candidate = sentence_a if len(sentence_a) > len(sentence_b) else sentence_b

        # remove one token from either end in equal probabilities

        if random.random() < 0.5:

            candidate.pop(0)

        else:

            candidate.pop()

 

 

if __name__ == “__main__”:

    print(time.time(), “started”)

    tokenizer = tokenizers.Tokenizer.from_file(TOKENIZER_PATH)

    print(time.time(), “loaded tokenizer”)

    docs = create_docs(PATH, NAME, tokenizer)

    print(time.time(), “created docs with %d documents” % len(docs))

    dataset = Dataset.from_generator(create_dataset, gen_kwargs={“docs”: docs, “tokenizer”: tokenizer})

    print(time.time(), “created dataset from generator”)

    # Save dataset to parquet file

    dataset.to_parquet(“wikitext-103_train_data.parquet”)

    print(time.time(), “saved dataset to parquet file”)

    # Load dataset from parquet file

    dataset = Dataset.from_parquet(“wikitext-103_train_data.parquet”, streaming=True)

    print(time.time(), “loaded dataset from parquet file”)

    # Print a few samples

    for i, sample in enumerate(dataset):

        print(i)

        print(sample)

        print()

        if i >= 3:

            break

    print(time.time(), “finished”)



Source_link

READ ALSO

MIT projects selected for funding under US Department of Energy’s Genesis Mission | MIT News

You Didn’t Get the AI Model You Paid For

Related Posts

MIT projects selected for funding under US Department of Energy’s Genesis Mission | MIT News
Al, Analytics and Automation

MIT projects selected for funding under US Department of Energy’s Genesis Mission | MIT News

July 23, 2026
You Didn’t Get the AI Model You Paid For
Al, Analytics and Automation

You Didn’t Get the AI Model You Paid For

July 23, 2026
Anthropic Releases Claude Security Plugin for Claude Code in Beta: A Multi-Agent Vulnerability Scanner That Runs in Your Terminal
Al, Analytics and Automation

Anthropic Releases Claude Security Plugin for Claude Code in Beta: A Multi-Agent Vulnerability Scanner That Runs in Your Terminal

July 23, 2026
Professor Emeritus Dimitri Bertsekas, influential computer scientist and prolific author, dies at 83 | MIT News
Al, Analytics and Automation

Professor Emeritus Dimitri Bertsekas, influential computer scientist and prolific author, dies at 83 | MIT News

July 22, 2026
Unsloth vs Axolotl vs TRL vs LLaMA-Factory: A Fine-Tuning Framework Comparison on Speed, VRAM, and Multi-GPU
Al, Analytics and Automation

Unsloth vs Axolotl vs TRL vs LLaMA-Factory: A Fine-Tuning Framework Comparison on Speed, VRAM, and Multi-GPU

July 22, 2026
FDA-Ready Medical Image Annotation Guide for Healthcare AI
Al, Analytics and Automation

FDA-Ready Medical Image Annotation Guide for Healthcare AI

July 22, 2026
Next Post
What Type Lawyer Makes The Most Money?

What Type Lawyer Makes The Most Money?

POPULAR NEWS

Trump ends trade talks with Canada over a digital services tax

Trump ends trade talks with Canada over a digital services tax

June 28, 2025
15 Trending Songs on TikTok in 2025 (+ How to Use Them)

15 Trending Songs on TikTok in 2025 (+ How to Use Them)

June 18, 2025
Communication Effectiveness Skills For Business Leaders

Communication Effectiveness Skills For Business Leaders

June 10, 2025
App Development Cost in Singapore: Pricing Breakdown & Insights

App Development Cost in Singapore: Pricing Breakdown & Insights

June 22, 2025
Comparing the Top 7 Large Language Models LLMs/Systems for Coding in 2025

Comparing the Top 7 Large Language Models LLMs/Systems for Coding in 2025

November 4, 2025

EDITOR'S PICK

I Evaluated G2 Reviews for 10 Free Form Builder Tools

I Evaluated G2 Reviews for 10 Free Form Builder Tools

January 24, 2026
Microsoft Research Releases Webwright: A Terminal-Native Web Agent Framework That Scores 60.1% on Odysseys, Up from Base GPT-5.4’s 33.5%

Microsoft Research Releases Webwright: A Terminal-Native Web Agent Framework That Scores 60.1% on Odysseys, Up from Base GPT-5.4’s 33.5%

May 24, 2026
The US gets the worst phones

The US gets the worst phones

April 25, 2026
OpenAI launches company knowledge in ChatGPT, letting you access your firm's data from Google Drive, Slack, GitHub

OpenAI launches company knowledge in ChatGPT, letting you access your firm's data from Google Drive, Slack, GitHub

October 24, 2025

About

We bring you the best Premium WordPress Themes that perfect for news, magazine, personal blog, etc. Check our landing page for details.

Follow us

Categories

  • Account Based Marketing
  • Ad Management
  • Al, Analytics and Automation
  • Brand Management
  • Channel Marketing
  • Digital Marketing
  • Direct Marketing
  • Event Management
  • Google Marketing
  • Marketing Attribution and Consulting
  • Marketing Automation
  • Mobile Marketing
  • PR Solutions
  • Social Media Management
  • Technology And Software
  • Uncategorized

Recent Posts

  • What is Agentic Search & How Does it Change SEO
  • Why Marketers Are Still Guessing What’s Driving Pipeline
  • Paycom Leans into Experiential to Make a Loyalty Play at SHRM26
  • EU fines Google $1 billion for favoritism in Search, Play Store
  • About Us
  • Disclaimer
  • Contact Us
  • Privacy Policy
No Result
View All Result
  • Technology And Software
    • Account Based Marketing
    • Channel Marketing
    • Marketing Automation
      • Al, Analytics and Automation
      • Ad Management
  • Digital Marketing
    • Social Media Management
    • Google Marketing
  • Direct Marketing
    • Brand Management
    • Marketing Attribution and Consulting
  • Mobile Marketing
  • Event Management
  • PR Solutions