Datasets:

vela-research
/

VCBench

pretty_name: VCBench (venture capital benchmark)
language:
  - en
license: other
license_name: vcbench-research-terms
license_link: https://vcbench.com
task_categories:
  - tabular-classification
  - text-classification
tags:
  - venture-capital
  - founder-success-prediction
  - startup-success-prediction
  - benchmark
  - leaderboard
  - finance
size_categories:
  - 1K<n<10K
extra_gated_prompt: >-
  VCBench is available on request for research use and may not be distributed or
  republished, so it stays out of the training corpus of future language models.
  Request access at https://vcbench.com (Request the dataset). This repository
  holds the dataset card only; the files are delivered by email after approval.

VCBench: the first benchmark for venture capital

Website and leaderboard: https://vcbench.com · Paper: arXiv:2509.14448 · Contact: benchmark@vela.partners

VCBench ranks LLMs, AI-native venture capital methods, classical ML and human investors on predicting which startup founders will succeed. It was created by the University of Oxford and Vela Research, the research arm of Vela Partners, an AI-native, quantitative venture capital firm in San Francisco.

Not to be confused with the vision-language math benchmark, the video benchmark or the genomics tool that also use the name VCBench.

The data

9,000 anonymized founder profiles collected from LinkedIn and Crunchbase; 810 (9%) labeled successful.
Success: the founder's company exited or IPO'd above a $500M valuation, or raised more than $500M.
A multi-stage pipeline standardizes, filters, enriches and anonymizes the profiles. Adversarial testing showed more than 90% reduction in re-identification risk while preserving predictive features (education quality, job history, industry).
Models are scored on a private test set of 4,500 founders (labels withheld), reported as the mean over three sequential folds.

The metric

F0.5, which weights precision above recall, because a bad bet costs an investor more than a missed one. Every entry also reports precision, recall and the API cost of scoring 1,000 founders at list prices.

Human baselines

Normalized to the dataset's 9% base rate, tier-1 VCs reach F0.5 10.7 and Y Combinator 8.6. The current leaderboard is at https://vcbench.com, with one page per model at https://vcbench.com/models.

Access

This repository is a dataset card. The files are not hosted here. Request the dataset at https://vcbench.com for research use. Each requester agrees not to distribute or republish it. To submit a model, email benchmark@vela.partners with predictions for the private test set.

Citation

@misc{chen2025vcbench,
  title={VCBench: Benchmarking LLMs in Venture Capital},
  author={Rick Chen and Joseph Ternasky and Afriyie Samuel Kwesi and Ben Griffin and Aaron Ontoyin Yin and Zakari Salifu and Kelvin Amoaba and Xianling Mu and Fuat Alican and Yigit Ihlamur},
  year={2025},
  eprint={2509.14448},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2509.14448}
}

Think-Reason-Learn: open-source LLM-native machine learning library implementing Policy Induction, Random Rule Forest, Reasoned Rule Mining and GPTree, several of the leaderboard's top entries.
Research using VCBench: papers and submissions by other groups.
Machine-readable summary: https://vcbench.com/llms.txt