PrimeBench: Practical Real-World Industry and Multi-Domain Evaluation Benchmark

An open benchmark for evaluating evaluators: 400 paired responses, deliberately edited along conflicting criteria, that test whether a reward model or LLM judge can tell a genuinely better answer from a plausible worse one.

Built and maintained by Composo.

Overview

Problem: Current evaluation datasets focus on overall quality (e.g., RewardBench) or single aspects (e.g., Anthropic HH-RLHF). However, these datasets are not practical for real-world use cases, which are more complex, domain-specific, and business-oriented. These existing datasets are insufficient for evaluating models against multiple specific criteria such as "Reward responses that demonstrate creativity." in real world scenarios.

Solution: PrimeBench provides paired responses edited along conflicting criteria (Comprehensive/Concise, Hallucinated/Factual, Technical/Simple) to test whether evaluation models can distinguish nuanced differences.

What's in it

400 examples across six criteria, drawn evenly from four real-world source datasets:

Source	Domain	Examples
FinQA	Financial question answering	100
XSum	News summarisation	100
PubMed	Biomedical literature	100
TechQA	Technical support	100

Dataset Design

1. Source: Real-world datasets (FinQA, XSUM, PubMed, TechQA) with questions and corresponding context
2. Baseline: Generates initial answers using the original context
3. Criteria Pairs:
   • Comprehensive vs Concise
   • Hallucinated vs Factual
   • Technical vs Simple
4. Modification: Create two responses that align with the given criteria pair.
5. Evaluation: Score both chosen and rejected answers using LLM judges and Composo models.
6. Comparison: If the score for the chosen answer is higher than the rejected answer, this data point is considered a pass; otherwise, it's a fail. The pass rate across all data points represents the performance of the evaluation approach.

Quick Start

# Clone and install
git clone https://github.com/composo-ai/PRIME
cd PRIME
pip install -r requirements.txt

# Configure (copy and edit with your API keys)
# For Composo API key, please email contact@composo.ai
cp config.example.json config.json

# Run evaluation
python Scripts/evaluate.py

# Show results
python Scripts/show_results.py

The dataset ships with this repo at data/dataset.json, and is also published on the Hugging Face Hub if you'd rather load it directly:

from datasets import load_dataset

dataset = load_dataset("ComposoAI/PrimeBench")

Data Format

{
    "prompt": "The input containing user question and context",
    "criterion": "The evaluation criteria being tested",
    "chosen": "The response that should score higher according to the criterion",
    "rejected": "The response that should score lower according to the criterion",
    "datasource": "The source dataset (e.g., FINQA, XSUM, PubMed, TechQA)"
}

Related work

• Composo Research - our papers, benchmarks and findings on evaluating AI
• On Cost-Effective LLM-as-a-Judge Improvement Techniques - what actually improves LLM judge accuracy on RewardBench 2, and what doesn't (code)
• Composo Align achieves state-of-the-art performance in evals - how PrimeBench was built and what we measured with it

Licence

Apache License 2.0 - see LICENSE. The same terms apply to the copy of the dataset published on Hugging Face.

Citation

@misc{primebench2025,
  title  = {PrimeBench: Practical Real-World Industry and Multi-Domain Evaluation Benchmark},
  author = {{Composo AI}},
  year   = {2025},
  url    = {https://github.com/composo-ai/PRIME}
}