lemon07r/pile-calibration-v3

Pile Calibration v3

A high-quality calibration dataset for LLM quantization, combining NeelNanda/pile-10k with bartowski's v3 imatrix calibration data.

Dataset Summary

Metric	Value
Total samples	9,767
Pile-10k samples	9,599
Bartowski v3 samples	168
Languages	English (primary), German, Spanish, French, Italian, Swedish, Russian, Arabic, Chinese

What's Included

This dataset merges two sources:

1. NeelNanda/pile-10k (filtered): A diverse subset of The Pile, filtered for quality
2. bartowski's calibration_datav3.txt: Curated multilingual calibration data with code, scientific text, dialogue, and more

The bartowski samples are randomly distributed throughout the dataset (not appended at the end) for better calibration coverage.

Quality Filtering Applied

The pile-10k source was filtered to remove low-quality samples. Starting from 10,000 samples:

Filter	Samples Removed	Description
Min length (< 200 chars)	226	Very short, low-content samples
Encoding corruption	1	Mojibake, replacement characters (�)
Line repetition	55	Same line repeated 5+ times, >25% of content
Low alphabetic ratio	33	Less than 25% alphabetic characters
Tag soup	80	Excessive HTML/XML markup with little actual content
Excessive whitespace	6	More than 60% whitespace
Total filtered	401
Samples kept	9,599

Examples of Filtered Content

Encoding corruption:

�I�v�V�����ƃv���[���тɂ���  H-RANDOM �� S-RANDOM...

Line repetition:

add your own caption
add your own caption
add your own caption
(repeated 11 times)

Low alphabetic ratio (math dumps):

Solve 0 = -6*x - 27*p + 276, 73*p - 72*p = 4*x - 146 for x.
37
Solve 6*n + 2828 = 5*z + 3287...

Tag soup:

<?xml version="1.0" encoding="UTF-8"?>
<segment><name>PD1</name><description>Patient...

Bartowski v3 Processing

The bartowski calibration data was processed using semantic chunking to detect document boundaries and prevent bleed-in (mixing of unrelated content). The algorithm detects:

• Language switches (English ↔ Russian, Spanish, German, etc.)
• Content type transitions (prose ↔ code, dialogue, math)
• Document start patterns (copyright headers, patents, story beginnings)
• Wikipedia end markers (Category: lines)
• Vocabulary shifts between paragraphs

This hopefully produced 168 semantically coherent samples, I couldn't detect any bleed-in with my quick preliminary analysis.

Usage

With Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("lemon07r/pile-calibration-v3", split="train")
print(f"Samples: {len(dataset)}")
print(dataset[0]["text"][:200])

With AutoRound

from auto_round import AutoRound

# Load calibration data
calib_dataset = load_dataset("lemon07r/pile-calibration-v3", split="train")

# Use with AutoRound
autoround = AutoRound(
    model,
    tokenizer,
    dataset=calib_dataset,
    # ... other parameters
)

For imatrix Calibration

Download the .txt file for use with llama.cpp imatrix:

# Download the text file
wget https://huggingface.co/datasets/lemon07r/pile-calibration-v3/resolve/main/pile-calibration-v3.txt

# Generate imatrix
./llama-imatrix -m model.gguf -f pile-calibration-v3.txt -o imatrix.dat

Files

File	Description
data-00000-of-00001.arrow	HuggingFace dataset format
pile-calibration-v3.jsonl	JSONL format for AutoRound
pile-calibration-v3.txt	Plain text for imatrix calibration

License

MIT License

Acknowledgments

• NeelNanda/pile-10k - Original pile subset
• bartowski1182 - v3 imatrix calibration data