Genμ 2.0
Frequently asked questions
Questions answered upfront
This FAQ covers the 2026 edition of Genμ. It clarifies how the legacy benchmark fits into the new identity-centric tracks, what data is used, and how to submit.
FAQ
Is the original Genμ challenge still included?
Yes. The earlier concept-unlearning challenge remains Task 3, while Tasks 1 and 2 cover face and speech identity unlearning.
Are face and speech participants restricted to one base model?
No. Tasks 1 and 2 are model-agnostic. Teams may use any suitable model, provided the exact model/checkpoint is declared and the submitted repository reproduces the outputs. Suggested Hugging Face checkpoints are listed on the baselines page.
What are the public validation identities?
The released CelebA and LibriTTS forget/retain groups are listed on the Face and Speech sections of the evaluation page and are also downloadable as CSV and JSON.
How do teams submit?
There are three track-specific Google Forms. Each submission must include a shareable drive link to the final weights, a GitHub repository containing the complete code, and detailed setup and testing instructions. Teams entering multiple tracks submit one form per track.
Where are submissions and rankings handled?
Submissions are collected through the three track-specific Google Forms. We evaluate the submitted artifacts directly and update the leaderboard on this website as verified results become available.
Can teams submit to more than one track?
Yes. Submissions to one, two, or all three tracks are welcome.
What matters most in evaluation?
For the Face and Speech tracks, leaderboard ranking is based only on ERB, which balances erasure accuracy and retain accuracy. Geometry, attributes, prosody, quality, and intelligibility are reported separately as supporting analysis.
What are the challenge dates?
The final submission deadline is 10 July 2026. Evaluation and final results declaration are scheduled for 25 July 2026.
Where are the submission forms?
The three form buttons are available on the Call for Submissions page. Until the final URLs are published, the buttons are marked “link pending.”
What is the Face/Speech leaderboard formula?
For both identity tracks, \(EA=100-FA\), and \(ERB=2(EA)(RA)/(EA+RA)\). ERB is the only value used to determine leaderboard rank. Track-specific preservation, quality, and intelligibility measurements are reported separately.
Can a team revise its submission?
Yes, until 10 July 2026. Submit the same track form again and mark it as a revision; the latest timestamped complete entry replaces earlier versions. After the deadline, checkpoints and substantive code cannot be replaced.
Are there publication opportunities?
Teams may submit their own method paper to the U&ME Workshop through regular review by 9 July 2026. Separately, the two highest-ranked eligible teams will be invited to co-author the Genμ 2.0 competition report, provided they also exceed the relevant official track baseline.