Discord

Checklist EUROAI (CEOAI) 2026 Contest Day 1 · Task 1

Faithful Explanations

Produce token-level rationales for a frozen DistilBERT sentiment classifier that satisfy a faithfulness probe, in three differently scored variants.

  • NLP
  • Explainability / rationale extraction

The task

A court-ordered audit concerns a frozen distilbert-base-uncased-finetuned-sst-2-english sentiment classifier. The contestant receives 100 pre-tokenised SST-2-style sentences with the model's predicted label and confidence, and must give rationales: subsets of real token positions (never [CLS] or [SEP]) of size at most min(12, floor(0.30 × T_real)).

A rationale passes the faithfulness probe (from ERASER) when the margin with all other tokens replaced by [MASK] keeps at least half the original margin (sufficiency) and masking the rationale removes at least 0.3 of it (comprehensiveness). Subtask 1 (Witness, 30 pts) asks for one rationale per sentence, compared with the smallest passing set built by adding tokens in order of the provided occlusion deltas; Subtask 2 (Counter-Argument, 35 pts) asks for two different rationales per sentence, scored by match quality times their Jaccard distance; Subtask 3 (Journalist, 35 pts) asks for one rationale compared with a hidden 2-of-3 consensus of attention rollout, occlusion and gradient × input.

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
Starter kit: sentences.json (100 sentences with token_ids, attention_mask, predicted_label, confidence), classifier_info.json, occlusion_pack.json (per-token occlusion deltas and reference margins), 25 practice sentences with labelled rationales and occlusion deltas, and the notebook 'Faithful Explanations (RAW).ipynb' with the probe and baselines.
You submit
One CSV with columns subtaskID, datapointID, answer, where answer is 1 if a token is in the rationale; datapointID = sid × 1000 + tok (subtasks 1 and 3) or sid × 2000 + slot × 1000 + tok (subtask 2).
Scoring
Per subtask, points = round(clamp((metric − floor)/(ceil − floor), 0, 1) × max). Subtask 1: pooled token-F1, floor 0.32, ceiling 0.90, 30 pts. Subtask 2: mean per-sentence pair score (rationale match × Jaccard distance), floor 0.25, ceiling 0.85, 35 pts. Subtask 3: token-F1 against the consensus set, floor 0.35, ceiling 0.90, 35 pts.
Rules
  • The classifier is frozen and may not be retrained.
  • Nothing uploaded is executed; the CSV is compared with hidden ground truth.
  • Submission limit 30, two final submissions (platform settings).
Format
Contest Day 1, 16 July 2026, 5 hours (10:30–15:30 local time; Nitro AI Judge 07:30–12:32 UTC), on-site at CREIC, Cluj-Napoca; three tasks, 100 points each.

Details

Year
2026, Cluj-Napoca, Romania (Babeș-Bolyai University)
Round
Contest Day 1 · Task 1
Language
English
License
Not stated by the source