unicorn-eval 2.3.2

Evaluation and adaption method for the UNICORN Challenge

🧪 UNICORN Evaluation Toolkit

Welcome to the official evaluation repository for the UNICORN Challenge — a benchmark for foundation models in pathology, radiology, and medical language processing. This repository provides:

• The official UNICORN Challenge evaluation code
• A growing library of adaptors used to turn frozen features into predictions in vision tasks.

🚀 Challenge Overview

The UNICORN Challenge evaluates how well foundation models generalize across multiple modalities with minimal task-specific supervision:

• 🧠 Language and Vision-Language tasks: algorithm directly outputs predictions — no adaptor required
• 👁️ Vision tasks: algorithms outputs frozen features, these are passed through adaptors — lightweight models - to generate predictions.

We provide a few built-in adaptors, but you're highly encouraged to propose your own!
We maintain the full list of adaptors available on the Supported Adaptors page.

📦 Adaptors vs. Algorithms: What's the Difference?

In vision tasks, submissions consist of:

• A feature extractor (your algorithm)
• An adaptor (used to turn features into predictions)

You can experiment with different adaptors on top of the same algorithm without using up your submission slots.
Want to try a different adaptor? Email us using the provided template (see below) — we’ll run it for you on existing features.

In language and vision-language tasks, the algorithm outputs predictions directly, so no adaptor is needed.

🧩 Contributing a Custom Adaptor

Have a better idea for how to turn features into predictions?

You’re welcome to contribute a custom adaptor! Here's how:

1. Add your adaptor to src/unicorn_eval/adaptors/.
2. Inherit from one of the base adaptor classes in base.py.
3. Open a pull request with:
   • Your adaptor code
   • A short README.md that covers:
     • A clear description of your method
     • A list of tasks, or task types your method is designed for
   • A unique name (we will include your team name in the adaptor name to ensure you receive credit). When naming your method, please be as specific as possible — for example, indicate details like the number of layers or specific settings — so that related methods with different configurations can be distinctly named.
   • Any additional dependencies in a requirements.txt (details on adding new requirements below)

✅ Once accepted, your adaptor becomes selectable at submission time — and your team gets full recognition when it’s used!

💡 Keep in mind: we prioritize originality. If your adaptor is too similar to an existing one, it may not be accepted — so submit early and make it your own!

Implementation requirements for contributing a new adaptor

• Your adaptor method must be implemented as a standalone function, following the baseline template base.py
• It must complete within the allowed time limit of 1h
• It must run on CPU
• Submissions will be evaluated for correctness, efficiency, and compliance with the challenge policies
• 🚨 Important: Pre-trained adaptors are not allowed! Be original — you can use the few-shots, for example, for fitting or training your adaptor, but don’t rely on pre-trained solutions

Dependencies

• Each method must be able to run in the provided isolated environment
• Additional dependencies can be requested, but:
  • Approval of new dependencies is not guaranteed, dependencies will be evaluated based on compatibility with other packages
  • Organizers reserve the right to modify the list of dependencies over time, though we aim to maintain compatibility with existing adaptors
  • When specifying dependencies, use the least restrictive version (e.g., package>=1.0.0) to ensure flexibility

💬 Teams are encouraged to share ideas and discuss approaches on the Grand Challenge forum. Support and Q&A will also be available through the forum.

📤 Requesting New Adaptor Runs

You can request us to apply additional adaptors to your existing vision submission without impacting your submission limit.

📧 Submission Instructions

1. Go to your submission URL: https://unicorn.grand-challenge.org/evaluation/<leaderboard-specific-number>/submissions/<your-submission-id>/
   (Use only this format — not other links)

2. For each submission that you want to rerun with a new adaptor, specify:

   • The full submission link
     Example: https://unicorn.grand-challenge.org/evaluation/30/submissions/bc9b9fe2-1f8d-4b9e-af7b-0edb87b127a4/
   • The new adaptor(s) you want to apply (chosen from the Supported Adaptors).
     ⚠️ Responsible use: You’re welcome to submit additional adaptor run requests over time. However, to ensure fair access for all participants, we ask that each request remains targeted and minimal (e.g., max 2 adaptors per leaderboard per request). Bulk or unfocused requests may be deprioritized.

3. Email your request to [e-mail] containing the following template:

Submission: https://unicorn.grand-challenge.org/evaluation/<leaderboard-specific-number>/submissions/your-submission-id/
Adaptors:
- teamname_adaptorX_v1
- teamname_adaptorY_v2

[Repeat for other submissions if needed]

Summary

Modality	What You Submit	Are Adaptors Used?	Submission Limit Applies To
👁️ Vision	Algorithm (feature extractor) + Adaptor	✅ Yes	Algorithm only
🧠 Language	Algorithm (predictive)	❌ No	Algorithm
🧠 Vision-Language	Algorithm (predictive)	❌ No	Algorithm