license: apache-2.0
task_categories:
- text-generation
language:
- en
size_categories:
- 10K<n<100K
Dataset Card for pg19-test
This dataset is a curated subset of the PG-19 dataset (a large collection of classic books from Project Gutenberg), specifically processed to generate text samples with controlled token lengths—using the Llama3.1-8B-Instruct tokenizer for precise tokenization—for long-context language model evaluation and research. The dataset contains stratified text excerpts at six distinct token length targets (10240 to 61440 tokens) with consistent sample sizes per length, ensuring sentence-level coherence (truncation at natural sentence endings) and minimal text reuse.
Dataset Details
Key Features
- Length Stratification: Samples across 6 target token lengths: 10k, 20k, 30k, 40k, 50k, 60k tokens.Note that the actual token length of samples does not strictly adhere to the target values, and a fluctuation of several hundred tokens (either above or below the target) may occur.
- Coherent Truncation: Text is truncated at natural sentence endings (.!?) rather than arbitrary token positions to preserve readability and semantic integrity.
- Language(s) (NLP): English
- Size of the dataset: 54.7MB
Dataset Sources
- Homepage: https://huggingface.co/datasets/hcyy/pg19-test
- Paper: SpecPV: Improving Self-Speculative Decoding for Long-Context Generation via Partial Verification
Dataset Structure
This dataset consists of 120 text samples structured into two core fields: "length" (an integer representing the target token length, which falls into six tiers: 10k, 20k, 30k, 40k, 50k, 60k tokens) and "text" (a string of coherently truncated PG-19 excerpts ending at natural sentence boundaries); it has 100 samples per length tier, with actual token lengths fluctuating by several hundred tokens around the targets to prioritize coherence, and is available in both Parquet formats.
Dataset Creation
Curation Rationale
This dataset is used to evaluate the performance of the SpecPV algorithm.
Source Data
The dataset is directly derived from the PG-19 dataset (hosted at https://huggingface.co/datasets/emozilla/pg19/), a large collection of public-domain books published before 1919 (from Project Gutenberg).
BibTeX:
[More Information Needed]
APA:
[More Information Needed]