haesol-shin/manta-1m-seqlen-513-1024

MANTA-1M: Calibration Subset (Seq Len 513-1024)

<p align="center"> <img src="https://huggingface.co/datasets/LGAI-EXAONE/MANTA-1M/resolve/main/Manta.png" alt="Manta" width="50%"> </p>

Overview

This dataset is a length-specific subset of the **LGAI-EXAONE/MANTA-1M** dataset, curated specifically for Post-Training Quantization (PTQ) Calibration.

Following the insights from the MaCa (Matryoshka Calibration) paper, this dataset provides length-specific calibration samples to ensure that the quantization process accounts for the variable weight importance across different input scales. By focusing on the 513-1024 token range, this subset helps capture "Late Channel" activation patterns that are often underrepresented in longer, fixed-length calibration sets.

Note: This is part of a multi-scale calibration series. For longer contexts (which capture "Early Channel" patterns), please refer to the [MANTA-1M: Calibration Subset (Seq Len 1025-2048)](https://huggingface.co/datasets/haesol-shin/manta-1m-seqlen-1025-2048).

Dataset Details

• —Source: Filtered from **LGAI-EXAONE/MANTA-1M**.
• —Filtering Criteria:
• —Sequence Length: 513 to 1024 tokens (Total Context: Prompt + Completion)
• —Tokenizer: [`LGAI-EXAONE/EXAONE-4.0-1.2B`](https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-1.2B)
• —Purpose: Optimized for Multi-scale Calibration, allowing models to maintain accuracy across various context lengths by providing a more stable and representative Hessian estimate.

Domain Diversity

This subset preserves the diverse domain distribution of the original MANTA-1M. (Percentages represent the original distribution and may vary slightly due to length filtering).

Usage

from datasets import load_dataset

# Load the calibration subset
dataset = load_dataset("haesol-shin/manta-1m-seqlen-513-1024", split="train")

# Shuffle and select random samples
calibration_set = dataset.shuffle(seed=42).select(range(128))

References & License

[1] Methodology Inspiration (MaCa): If you use this dataset for quantization research, please cite the MaCa paper:

@article{son2026maca,
  title={On the Importance of a Multi-Scale Calibration for Quantization},
  author={Son, Seungwoo and Seong, Ingyu and Kim, Junhan and Jang, Hyemi and Jeon, Yongkweon},
  journal={arXiv preprint arXiv:2602.07465},
  year={2026}
}

[2] Data License: This dataset is released under the CC-BY-NC-4.0 License.

• —Non-Commercial Use Only.
• —Attribution: Based on the MANTA-1M dataset by LG AI Research.

Contact

For inquiries regarding the original data, contact: [contact_us@lgresearch.ai](mailto:contact_us@lgresearch.ai)