Skip to content

Integrations and DLC#

Optional extras patch popular libraries to make corruption frictionless. This page is a quick reference catalog—for detailed tutorials and examples, see Pipeline Workflows.

Hugging Face Datasets (hf extra)#

  • Use glitchlings.dlc.huggingface.GlitchedDataset to wrap a dataset and corrupt one or more explicit text columns.
  • Reuse gaggle seeds to reproduce corrupted datasets across machines.
from glitchlings.dlc.huggingface import GlitchedDataset

corrupted = GlitchedDataset(dataset, "typogre", column="text", seed=404)

PyTorch (torch extra)#

  • Use glitchlings.dlc.pytorch.GlitchedDataLoader to wrap a DataLoader.
  • Accepts glitchling names, instances, or gaggles; auto-inferring text columns from batches when columns is omitted.
from glitchlings.dlc.pytorch import GlitchedDataLoader

glitched = GlitchedDataLoader(loader, ["typogre", "mim1c"], seed=404)

Lightning (lightning extra)#

  • Use glitchlings.dlc.pytorch_lightning.GlitchedLightningDataModule to wrap a LightningDataModule. Specify the text column(s) to corrupt.
  • Designed for evaluation corruption with minimal boilerplate.
from glitchlings.dlc.pytorch_lightning import GlitchedLightningDataModule

glitched = GlitchedLightningDataModule(datamodule, "typogre", column="text", seed=404)

LangChain (langchain extra)#

  • Use glitchlings.dlc.langchain.GlitchedRunnable to wrap LCEL runnables and glitch inputs (optionally outputs) without modifying the chain.
  • Columns/fields are inferred from the first payload when omitted; pass input_columns/output_columns for explicit control.
from glitchlings.dlc.langchain import GlitchedRunnable
from glitchlings import Typogre

glitched = GlitchedRunnable(chain, Typogre(rate=0.01), glitch_output=True, seed=404)
response = glitched.invoke({"question": "Who guards the guardians?"})

NVIDIA NeMo DataDesigner (nemo extra)#

  • Install the nemo extra for glitchlings.dlc.nemo, or install the standalone plugin package glitchlings-nemo for automatic DataDesigner discovery.
  • GlitchlingColumnConfig defines a text corruption column generator.
  • Accepts glitchling names, specs with parameters, lists, or YAML config paths.
  • Use source_column to corrupt a different column than the output.
from data_designer import DataDesignerConfigBuilder
from glitchlings.dlc.nemo import GlitchlingColumnConfig

builder = DataDesignerConfigBuilder()
builder.add_column(
    GlitchlingColumnConfig(
        name="corrupted_prompt",
        source_column="prompt",
        glitchlings=["Typogre(rate=0.02)", "Mim1c(rate=0.01)"],
        seed=404,
    )
)

Standalone Usage (without DataDesigner)#

For direct DataFrame corruption without the full DataDesigner infrastructure:

import pandas as pd
from glitchlings.dlc.nemo import corrupt_dataframe

df = pd.DataFrame({"text": ["Hello world", "Test input"]})
result = corrupt_dataframe(df, "typogre", column="text", seed=42)

Flexible Glitchling Specifications#

The plugin accepts multiple specification formats:

# Pre-constructed Gaggle
from glitchlings import Gaggle, Typogre, Mim1c
gaggle = Gaggle([Typogre(rate=0.02), Mim1c(rate=0.01)], seed=404)

# Auggie fluent builder
from glitchlings import Auggie
auggie = Auggie(seed=404).typo(rate=0.02).confusable(rate=0.01)

# String specification
glitchlings = "Typogre(rate=0.02)"

# List of specifications
glitchlings = ["Typogre(rate=0.02)", "Mim1c(rate=0.01)"]

# YAML config path
glitchlings = "configs/chaos.yaml"

Prime Intellect (prime extra)#

  • Install the prime extra for glitchlings.dlc.prime.
  • load_environment wraps verifiers.load_environment and injects glitchlings into benchmarks.
  • echo_chamber bootstraps text-cleaning challenges directly from Hugging Face datasets.
  • Pass seed= to keep corrupted environments deterministic.

Project Gutenberg (gutenberg extra)#

  • Install the gutenberg extra for glitchlings.dlc.gutenberg.
  • GlitchenbergAPI wraps the py-gutenberg GutenbergAPI and corrupts book text on fetch.
  • Accepts glitchling names, instances, or gaggles; seeds for deterministic corruption.
  • Original titles are preserved in original_title for comparison.
  • Use get_text() to fetch and corrupt the full book content.
from glitchlings.dlc.gutenberg import GlitchenbergAPI

api = GlitchenbergAPI("typogre", seed=42)
book = api.get_book(1342)  # Pride and Prejudice

# Access corrupted and original titles
print(book.title)           # Corrupted title
print(book.original_title)  # "Pride and Prejudice"

# Fetch and corrupt the full text content
full_text = book.get_text()
print(full_text[:100])  # First 100 chars of corrupted text

Custom Gutendex Instance#

By default, GlitchenbergAPI uses the public Gutendex instance. For production use or high-volume requests, you can specify a custom instance URL:

from glitchlings.dlc.gutenberg import DEFAULT_GUTENDEX_URL, GlitchenbergAPI

# Use default public instance
api = GlitchenbergAPI("typogre")

# Or specify a custom/self-hosted Gutendex instance
api = GlitchenbergAPI("typogre", instance_url="https://my-gutendex.example.com")

Batch Processing#

For batch corruption of books fetched from other sources:

from glitchlings.dlc.gutenberg import GlitchenbergAPI

api = GlitchenbergAPI(["typogre", "mim1c"], seed=42)

# Corrupt multiple books at once
books = api.get_books_by_search("shakespeare")
for book in books:
    print(f"{book.original_title} → {book.title}")

Installing extras#

pip install 'glitchlings[hf]'          # datasets
pip install 'glitchlings[torch]'       # PyTorch DataLoader
pip install 'glitchlings[lightning]'   # Lightning DataModule
pip install 'glitchlings[langchain]'   # LangChain runnables
pip install 'glitchlings[nemo]'        # NeMo DataDesigner
pip install 'glitchlings[prime]'       # Prime Intellect DLC
pip install 'glitchlings[gutenberg]'   # Project Gutenberg
pip install 'glitchlings[all]'         # everything

# Alternatively, install the standalone NeMo plugin for auto-discovery:
pip install glitchlings-nemo