Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
33.6k
187k
source
stringclasses
7 values
sequence_avg_median_lookback
float64
25.5
2.6k
sequence_avg_max_lookback
float64
423
7.02k
\section{Introduction} Wireless communication systems are challenged to fulfill novel demanding requirements and provide wider coverage for all users, whose number is constantly increasing. For that reason, there is a need to introduce novel techniques, and/or to combine existing ones. With regard to radio-frequency (...
arxiv
2,602.905762
5,204.026855
\subsection{Detailed Dataset Statistics} Table \ref{tab:statistics} demonstrates the statistic information of different splits of our proposed \textit{AntCritic}. The upper part demonstrates the quantitative characteristics of texts, and the lower part shows the percentage of segments that contains the specified HTML t...
arxiv
2,479.429932
5,111.155273
"\\section{Introduction}\n\nA considerable amount of evidence suggests that $\\sim$84\\% of the tota(...TRUNCATED)
arxiv
2,458.992432
5,342.897461
"\\section{Introduction}\n\nChinese may use different Chinese characters up to 20,000 so that it is (...TRUNCATED)
arxiv
2,370.889893
4,947.111328
"\\section{Introduction}\nPolitical conversations between everyday people form the foundation of a h(...TRUNCATED)
arxiv
2,361.234619
4,779.759766
"\\section{Introduction}\r\n\r\nSystems of two-dimensional (2D) electrons close to van Hove (vH)\r\n(...TRUNCATED)
arxiv
2,326.185791
5,314.365723
"\\section{INTRODUCTION}\n\nHyperledger Fabric is a popular permissionless system that allows the de(...TRUNCATED)
arxiv
2,286.673828
4,869.537109
"\\section*{Introduction}\nOrdered dynamical phases of motile organisms are ubiquitous in nature acr(...TRUNCATED)
arxiv
2,203.792969
5,109.62793
"\\section{Introduction}\n\\label{Introduction}\n\n\\subsection{Time and Clocks}\n\\label{TimeObserv(...TRUNCATED)
arxiv
2,183.296143
5,366.569824
"\\section{Introduction}\n\tStructural analysis represents the set of theories and methods which all(...TRUNCATED)
arxiv
2,171.375488
4,668.36377
End of preview. Expand in Data Studio

Attention-Labeled 16K Documents

Documents labeled by attention lookback distance at 16K context length, used for attention-based data filtering for long-context language model training.

Overview

Each document was processed through Qwen2.5-Coder-7B-Instruct at 16,384 tokens, and the attention pattern at layer 27 was analyzed to measure how far back the model looks when processing the text. Documents where the model attends to distant tokens (high lookback) are labeled positive/filtered — these are hypothesized to be better training data for teaching long-context capabilities.

Files

File Rows Description
filtered_128.parquet 1,164 Positive — high attention lookback (threshold ≥ 128 tokens), diverse sources
filtered_300.parquet 168 Positive — stricter threshold (≥ 300 tokens)
negative_128.parquet 1,085 Negative — low attention lookback (below 128-token threshold)
negative_300.parquet 168 Negative — matched count for the 300-token threshold
random_match128.parquet 1,164 Random baseline — count-matched to filtered_128
random_match300.parquet 168 Random baseline — count-matched to filtered_300

Fields

Field Type Description
text string Full document text (all files)
source string Data source domain (filtered/random files only)
sequence_avg_median_lookback float Median attention lookback distance averaged across sequence positions (filtered/random files only)
sequence_avg_max_lookback float Max attention lookback distance averaged across sequence positions (filtered/random files only)

Source Domains

Documents were drawn from a pool of long-form documents (≥16K tokens) from multiple domains:

Domain Description
code Source code from The Stack
arxiv Academic papers
web FineWeb web text
legal Legal documents (MultiLexSum)
books Books (PG-19)
encyclopedia Wikipedia
government Government reports

Labeling Details

  • Model: Qwen2.5-Coder-7B-Instruct
  • Context length: 16,384 tokens
  • Attention layer: 27 (final layer)
  • Metric: Sequence-level average median lookback distance
  • Block size: 4096 query × 4096 key
  • Filler tokens: 128 (added to probe long-range attention)

Usage

from datasets import load_dataset

# Load all splits
ds = load_dataset("KevinDavidHayes/labeled-16k-attention")

# Or load specific files
filtered = load_dataset("KevinDavidHayes/labeled-16k-attention", data_files="data/filtered_128.parquet")

Citation

If you use this dataset, please cite our work on attention-based data filtering for long-context training.

Downloads last month
46