DeepInfra/variable-length-embedding-benchmark

Variable Length Embedding Benchmark (VLEB)

Dataset Summary

VLEB (Variable Length Embedding Benchmark) is a specialized dataset designed to evaluate the performance, latency, and stability of embedding models and rerankers across a wide spectrum of context lengths.

Unlike standard datasets that focus on short passages, VLEB provides a balanced distribution of text ranging from standard RAG chunks to maximum-context documents (up to 32k tokens). It is constructed from wikitext-103-raw-v1 using a smart-clipping strategy that preserves semantic integrity without splitting words.

This benchmark is essential for:

• Length Generalization: Testing if models maintain semantic understanding as context grows.
• RAG Profiling: Measuring encoding latency and memory usage at different bins.
• "Lost-in-the-Middle" Analysis: Evaluating retrieval degradation in long-context windows.

Data Structure

The dataset consists of 8,000 samples, strictly balanced across 4 length categories (2,000 samples per bin). Token counts are calculated using the Qwen/Qwen2.5-7B-Instruct tokenizer.

Category	Token Range (Qwen)	Typical Use Case
Short	512 - 2,048	Standard RAG chunks, abstracts, news snippets.
Medium	2,048 - 8,192	Full articles, technical reports, single-file code.
Long	8,192 - 16,384	Multiple papers, book chapters, long legal contracts.
Very Long	16,384 - 32,000	Entire books, massive documentation, stress testing context limits.

Usage

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("ovuruska/variable-length-embedding-bench")

# Filter for specific length requirements
short_contexts = dataset.filter(lambda x: x['length_category'] == 'Short')
very_long_contexts = dataset.filter(lambda x: x['length_category'] == 'Very Long')

print(f"Sample Text ({very_long_contexts[0]['length_category']}):")
print(very_long_contexts[0]['text'][:200] + "...")

Construction Methodology

1. Source: The dataset is derived from the wikitext-103-raw-v1 corpus.
2. Stream Buffering: The raw text was processed as a continuous stream rather than isolated lines.
3. Smart Clipping: A buffer system accumulated tokens until a target length (randomly selected within bin ranges) was met. The text was then clipped at the exact token boundary and decoded back to string, ensuring no words are split and the text remains natural.
4. Validation: All samples were re-tokenized to ensure they strictly fall within their assigned bin limits.

Citation

If you use this dataset for benchmarking, please cite:

@misc{vleb_2026,
  author = {DeepInfra Engineering Team},
  title = {Variable Length Embedding Benchmark (VLEB)},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{[https://huggingface.co/datasets/DeepInfra/variable-length-embedding-benchmark](https://huggingface.co/datasets/DeepInfra/variable-length-embedding-benchmark)}}
}