Discord

Checklist AICC 2026 Round 9 · Task 3

Shuffled

Restore the original row order of the row-shuffled vision and text positional-embedding matrices of a CLIP ViT-B/16 model.

  • Vision
  • NLP
  • Multimodal
  • Model repair (recovering permutations of positional embeddings)
Read the official statement

The task

An offline copy of a vision-language model was damaged: its weights survived, but the two positional-embedding tables came back with their rows permuted — 196 rows for the image patches and 77 rows for the caption slots. A maintenance log preserves six of the original row indices.

The model is CLIP ViT-B/16. The vision table (excluding the unshuffled CLS row) has shape (196, 768), where row i originally held grid cell (i // 14, i % 14) of a 14 × 14 grid; the text table has shape (77, 512), where row i held sequence slot i. Independent random permutations were applied to the two tables. anchors.csv gives the true index of four vision rows and two text rows. One hundred matched image-caption pairs are supplied for local checking only.

For every shuffled row the contestant predicts its original index: 0–195 for vision rows and 0–76 for text rows, 273 rows in total, of which the six anchor rows are not scored.

Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.

At a glance

You get
clip/ (CLIP ViT-B/16 with both positional-embedding matrices row-shuffled), anchors.csv (row_id, position for 6 rows), data/ (100 matched image-caption pairs), sample_submission.csv.
You submit
submission.csv with columns row_id,position (row ids vision_i and text_i), one row per shuffled row in any order.
Scoring
Accuracy over the 267 scored rows; a row counts only if the predicted index equals the original index exactly. Baseline 0.02, reference solution 1.00.
Rules
  • Individual participation (maximum team size 1); at most 60 submissions per day.
  • AICC contest rules (stated on each Kaggle rules page, not enforceable): no use of LLMs for writing code or getting task ideas; no internet use other than official library documentation and the contest platform; no communication with anyone during the contest; clarifications only via the #clarification-requests channel on the AICC Discord server.
Format
AICC Round 9, online on Kaggle, 24 Jul 2026 18:05 UTC – 26 Jul 2026 18:00 UTC.

Notebook

Competition

Solution

Details

Year
2026, Online (Kaggle)
Round
Round 9 · Task 3
Language
English
License
Varies by task: Word Lookups — MIT; Buried Fault and Shuffled — CC BY-NC-ND 4.0 (Kaggle competition licences). Solutions repository: MIT., as stated by the source