Dataset Card: manipulation-resistant-prompts-96-96

Dataset Description

This dataset contains prompts with specified target word counts for both input prompts and target outputs, designed to test and evaluate language models across different length requirements. Word counts are defined as whitespace-separated tokens, providing a consistent and human-interpretable measure of text length.

These datasets are typically used in performance benchmarking of language models, where prompts are randomly sampled from the dataset to measure throughput, latency, and other performance metrics. The prompts consist of natural language text with coherent semantic content—they are not synthetically generated token sequences or random character strings, as such inputs would not accurately reflect real-world usage patterns and could lead to misleading performance measurements.

Why Words, Not Tokens?

This dataset uses word counts rather than token counts for several important reasons:

• Natural Human Thinking: Human communication, writing, and thinking occur in words, not tokens. When people specify requirements, they think in terms of "I need a 500-word essay" or "write a 100-word summary," not "I need 750 tokens."
• Universal Understanding: A word is a concrete, universally understood unit (defined as whitespace-separated text). Tokens, in contrast, are implementation-dependent units that vary significantly across tokenization algorithms. The same text can have substantially different token counts depending on the tokenizer used (e.g., tiktoken for GPT models, SentencePiece for Llama models, or proprietary tokenizers for Claude models).
• Real-World Applications: In practice, requirements are specified in words: academic papers have word limits, articles have word counts, summaries are requested in words. This dataset reflects how people actually work with text generation in real-world scenarios.
• Consistency Across Models: Word counts provide a stable, model-agnostic metric for comparing performance across different language models and their associated tokenizers, whereas token counts would require separate normalization for each tokenization scheme.
• Accessibility: Word counts are immediately understandable to anyone, regardless of their technical knowledge of tokenization algorithms or vocabulary sizes.
• Token Conversion: Words can easily be converted to tokens for any given model through the model's tokenizer, allowing benchmarks to obtain token-level metrics when needed while maintaining a universal word-based baseline for cross-model comparison.
• Gaming Resistance: Using words makes benchmarks more resilient against "gaming" of benchmarks through hardcoded optimizations. When benchmarks use fixed token counts, serving infrastructure can be optimized for those specific token bucket sizes, KV cache window sizes, or attention mechanism configurations. Word-based datasets prevent such targeted optimizations, making benchmarks less susceptible to manipulation and more likely to yield real-world performance metrics (e.g., tokens/second throughput) that are meaningful to actual users.
• Prompt Randomization: Random sampling of prompts from these datasets is essential for unbiased performance evaluation. The natural language content ensures that benchmarks reflect realistic inference workloads, as opposed to synthetically generated token sequences or random character strings that would not accurately represent actual model usage and could produce misleading performance measurements.

Dataset Summary

• Total Prompts: 782
• Created: 2025-11-08
• Target Prompt Length: 96 words
• Target Output Length: 96 words

Dataset Structure

The dataset contains the following columns:

• prompt: The text prompt (string)
• prompt_length: Target length for the prompt in words (integer)
• target_output_length: Target length for the output in words (integer)
• actual_words: Actual word count in the prompt (integer)

Column Statistics

Actual Words

Central Tendency:

• Mean: 102.66
• Median: 103.00
• Mode: 104

Dispersion:

• Standard Deviation: 1.82
• Variance: 3.30
• Range: 7 (98 to 105)
• Interquartile Range (IQR): 3.00

Quartiles:

• Q1 (25th percentile): 101.00
• Q2 (50th percentile / Median): 103.00
• Q3 (75th percentile): 104.00

Percentiles: 10th: 100.0, 25th: 101.0, 50th: 103.0, 75th: 104.0, 90th: 105.0, 95th: 105.0, 99th: 105.0

Actual Words Distribution

• Histogram: Peak: 160 prompts (20.5%) in range [104-104]
• Top bins: [104-104]: 160 (20.5%), [105-105]: 145 (18.5%), [103-103]: 135 (17.3%)
• Percentiles: 10th: 100.0, 25th: 101.0, 50th: 103.0, 75th: 104.0, 90th: 105.0, 95th: 105.0, 99th: 105.0

Actual vs Target Word Count Analysis

This compares the actual_words column (actual word count) with the prompt_length column (target word count).

• Mean Difference: 6.66 words
• Standard Deviation: 1.82 words
• Min Difference: 2 words
• Max Difference: 9 words
• Difference Percentiles: 10th: 4.0, 25th: 5.0, 50th: 7.0, 75th: 8.0, 90th: 9.0, 95th: 9.0, 99th: 9.0

Usage

from datasets import load_dataset

dataset = load_dataset("metrum-ai/manipulation-resistant-prompts-96-96")
print(dataset)

# Access the data
for example in dataset['train']:
    print(f"Prompt: {example['prompt'][:100]}...")
    print(f"Target prompt length: {example['prompt_length']}")
    print(f"Target output length: {example['target_output_length']}")
    print(f"Actual words: {example['actual_words']}")

Note on Token Conversion: The prompt word counts in this dataset can easily be converted to token counts depending on the tokenizer used for a given specific test. Simply tokenize the prompt text using your model's tokenizer to obtain the exact token count for that specific tokenization scheme.

License

This dataset is released under the MIT License.

MIT License

Copyright (c) 2025 Metrum AI

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Terms of Use

• Use: This dataset may be used for research, commercial, and non-commercial purposes.
• Attribution: Attribution to Metrum AI is required when using this dataset. Please cite the dataset using the citation provided below.
• Modification: You may modify, transform, and build upon this dataset.
• Distribution: You may distribute this dataset and any modifications you make to it, provided that attribution to Metrum AI is maintained.
• No Warranty: This dataset is provided "as is" without warranty of any kind, express or implied, including but not limited to the warranties of merchantability, fitness for a particular purpose, and noninfringement.
• Limitation of Liability: In no event shall Metrum AI be liable for any claim, damages, or other liability arising from the use of this dataset.

Citation

@dataset{manipulation_resistant_prompts_96_96,
  title = {manipulation-resistant-prompts-96-96},
  author = {Metrum AI},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/metrum-ai/manipulation-resistant-prompts-96-96}
}