Challenge

Where the Wild Features Are

Predict kindergarteners' end-of-year reading scores from literacy screening audio, and find which speech features carry signal beyond correctness scores

Tasks
Reading assessment
Data types
Audio
Tabular
$50,000 in prizes
8 weeks left model submissions
53 joined
Adapted from a photo by Guy Basabose on Unsplash

Solution write-up

In Phase 2, the solution write-up is where you explain what your model learned from children's speech. Prizes are awarded based on the write-ups, not on leaderboard ranking alone. Only the top 15 teams on the private leaderboard when Phase 1 closes will be invited to submit a write-up and are eligible for prizes.

The central question is: which acoustic features carry signal about end-of-year reading outcomes beyond the beginning-of-year screening scores (item scores and task ability estimates), and how much do they add? Strong write-ups answer this with rigorous analysis, show evidence that the signal comes from the child's speech, and are useful to literacy and speech researchers who may not be machine learning experts.

Timeline


  • January 19, 2027: Phase 1 submissions close and the leaderboard is frozen.
  • January 20, 2027: Invited teams are notified.
  • February 8, 2027, 23:59:59 UTC: Write-ups due.
  • February – March 2027: Judging and verification period. Finalists will submit code that reproduces their leaderboard submission. See the competition rules for code requirements.
  • Late March 2027: Winners announced.

Because the write-up depends on how you built and validated your model, plan for it during Phase 1. Keep track of your validation results and of which features you tested.

Additional data


After Phase 1 closes, invited teams will receive additional data to analyze their Phase 1 model in their write-ups. Using it is encouraged but not required. For example, it lets you check whether your results hold across subgroups, which judges score under Generalizability.

  • Test set labels
  • Test school_id and score_status
  • Each student's sex (where available) and year of data collection
  • Each clip's response time and item name, which identifies the specific prompt the child heard

Raw (non-anonymized) audio will not be provided. Your Phase 1 leaderboard submission is final.

Format requirements


  • PDF submission, uploaded on the "Write-up submissions" tab
  • Maximum of 4 pages, including figures and tables, not counting references
  • Page size: 8.5x11" with 1" margins
  • Font: minimum 11pt

Write-up template


Write-ups should include the following sections:

  1. Summary. A short overview of your approach and your most important findings.

  2. Key findings. Which acoustic features or feature groups carry signal beyond the screening scores, and how much they add. Key findings must come from the model you submitted in Phase 1. Analysis added after Phase 1 closes is allowed but must be reported separately as exploratory post-hoc analysis.

  3. Methodology. How you processed the audio, engineered or learned features, built and validated your model, and measured the contribution of acoustic features. Explain the reasoning behind important choices. Figures and tables are encouraged where they improve clarity.

  4. Interpretability. What your most important features mean, which reading or speech construct they relate to, what they suggest about children's speech and reading development, and how you showed which features or feature groups drive your model's predictions, for example with feature importance, ablation, or heat maps.

  5. Limitations and generalizability. Whether your findings hold across subgroups, such as outcome assessment, year of data collection, school, or sex; the main risks and open questions; and when your results should be trusted or treated with caution. Optionally, note what your findings suggest for literacy screening or follow-up research.

Evaluation criteria


A panel of subject matter experts will score write-ups on the following criteria.

Scientific insight (30%)

What the findings reveal about children's reading and speech

  • Which speech and acoustic features carry signal beyond, or separate from, the screening scores, and how much do they add?
  • Are the features novel, going beyond known measures such as response duration, speaking rate, or pauses?

Rigor and validity (20%)

Whether the results are rigorous on the data tested

  • Is there sound out-of-sample evidence for the findings?
  • Does the write-up show evidence that the signal comes from the child's speech, and check for possible confounds, such as adult speech from teachers or administrators, background noise, or room acoustics?
  • Does the write-up keep results for the model submitted in Phase 1 separate from new analysis done after the test labels were released? Features chosen after seeing the test labels are not tested out of sample, so report them as exploratory.

Model interpretability (20%)

Whether the model is transparent, and whether we can see what drives its predictions

  • Can the model's features be understood? Do they map to a reading or speech construct?
  • Does the write-up show which features or feature groups drive the predictions, for example with feature importance, ablation, or heat maps?

Generalizability (20%)

Whether the results hold beyond the data tested, across subgroups and settings

  • Do the findings hold across subgroups? For example, do error rates differ by outcome assessment, year of data collection, school, or sex?
  • Is the model parsimonious, avoiding unnecessary complexity?

Clarity (10%)

  • Are the ideas clear and well organized?

Write-ups are judged on the clarity of ideas and organization, not on English fluency. Grammar and phrasing will not be penalized.


If you have any questions, post in the competition forum or email k12-ai-infra@drivendata.org.