text-curation 1.0.0

Profile-based text curation pipelines for Hugging Face Datasets

text-curation

Profile-based, deterministic text curation pipelines for Hugging Face Datasets

---

Overview

text-curation is a Python library for building structured, profile-driven text curation pipelines designed to integrate naturally with Hugging Face Datasets.

It focuses on deterministic, inspectable, and conservative text transformations for preparing large-scale corpora used in NLP and LLM training workflows.

Rather than relying on ad-hoc cleanup scripts, text-curation promotes explicit, versioned curation profiles that make data decisions reproducible, auditable, and stable over time.

---

Design Principles

• Profile-driven pipelines
  Reusable, declarative profiles define how text is curated for a given domain (e.g. web, wiki, forums).

• Composable blocks
  Each transformation is implemented as an isolated block with a single, explicit responsibility.

• Deterministic and non-destructive
  All transformations are rule-based and conservative by default, prioritizing semantic preservation.

• Structure-aware
  Text is treated as structured data (paragraphs, lists, headers), not just as raw strings.

• Dataset-scale friendly
  Designed to run efficiently on large Hugging Face Datasets using .map.

---

Scope & Stability (v1.0.0)

As of v1.0.0, text-curation provides a stable core for structure-aware text curation of real-world, messy data.

The default behavior and semantics of the core blocks and profiles are considered stable and will not change without a major version bump.

The library intentionally focuses on deterministic preprocessing rather than semantic classification or machine-learning-based filtering.

---

Core Blocks (Stable)

The following blocks are part of the stable core in v1.0.0.

• Normalization
  Canonicalizes Unicode and typography (quotes, dashes, ellipses), removes control and zero-width characters, and normalizes whitespace.

• Formatting
  Reconstructs paragraphs, normalizes whitespace and punctuation spacing, and preserves relative indentation for structured content such as code blocks.

• Redaction
  Masks sensitive content such as emails, API tokens, and embedded credentials using explicit, non-destructive placeholders.

• Structure
  Detects headers, lists, repetition, and boilerplate indicators, emitting inspectable signals without mutating text.

• Filtering
  Applies conservative, signal-based removal of empty or low-value paragraphs.

• Deduplication
  Performs exact, paragraph-level deduplication using normalization-safe keys.

More aggressive semantic filtering, fuzzy deduplication, or heuristic cleanup are intentionally not enabled by default and may be introduced only via explicit opt-in profiles in future releases.

---

Non-Goals

text-curation intentionally does not attempt to:

• Perform semantic or topical classification
• Use machine learning or probabilistic heuristics
• Infer document quality or intent
• Preserve exact visual formatting of source text
• Aggressively remove all boilerplate or repetition by default

These constraints are by design and are critical to ensuring predictable, reproducible dataset preprocessing.

---

Installation

text-curation supports Python ≥ 3.9.

Install from PyPI:

pip install text-curation

Or install from source for development:

git clone https://github.com/Dhiraj309/text-curation.git
cd text-curation
pip install -e .

---

Quickstart

Curating a Hugging Face Dataset

from datasets import load_dataset
from text_curation import TextCurator

dataset = load_dataset(
    "allenai/c4",
    "en.noclean",
    split="train",
)

curator = TextCurator.from_profiles(profile_name="web_common_v1")

cleaned = dataset.map(
    curator,
    batched=True,
    num_proc=4,
)

The curator is a pure function: it takes a batch dictionary and returns a dictionary with the same schema.

---

Profiles

Profiles define which blocks are applied and in what order.

Example (conceptual):

web_common_v1 = [
    NormalizationBlock(),
    FormattingBlock(),
    RedactionBlock(),
    StructureBlock(),
    FilteringBlock(),
    DeduplicationBlock(),
]

Profiles are explicitly versioned to ensure reproducibility and long-term auditability of dataset preprocessing.

---

Designed For

• Web-scale datasets (C4-like, Common Crawl, scraped corpora)
• OCR- and PDF-derived text
• Forums, blogs, and user-generated content
• Dataset preprocessing prior to LLM training or evaluation

The library is intentionally model-agnostic and does not depend on tokenizers, embeddings, or classifiers.

---

Why text-curation?

• Cleaning text is not just normalization — structure and repetition matter
• Ad-hoc scripts do not scale or reproduce
• Dataset curation deserves the same rigor as model training
• Explicit pipelines make data decisions inspectable and debuggable

text-curation is designed to be the data-side analogue of model-definition libraries in the Hugging Face ecosystem.

---

When should you not use text-curation?

• If you only need a one-off regex cleanup
• If your data is already fully curated
• If you require ML-based content classification or scoring

---

Versioning & Compatibility

This project follows semantic versioning.

• 1.x releases guarantee stable default behavior
• Breaking changes require a major version bump
• Profiles are versioned explicitly (e.g. web_common_v1) to preserve reproducibility across releases

---

Contributing

Contributions are welcome.

When adding new blocks or profiles:

• Keep transformations deterministic
• Avoid destructive defaults
• Include clear before/after examples

See CONTRIBUTING.md for details.

---

License

Apache 2.0. See LICENSE.

---

Acknowledgements

Inspired by large-scale dataset curation practices in the Hugging Face ecosystem.