task_cards/real/affectivetext.md

SemEval AffectiveText emotions Task Card

• Task ID: affectivetext_emotions
• Subset: real
• Samples: 1000
• Simplex dimension: 6
• Predictor: Frozen zero-shot emotion scorer

Evaluation Role

Small-sample weak-structure counterexample with a frozen language-model predictor.

Target And Predictor

Normalized six-emotion gold scores from SemEval-2007 AffectiveText paired with a frozen zero-shot scorer cache.

Default Benchmark Settings

• Default score: Aitchison distance.
• Default stratification: Boundary bins on the predicted emotion mixture.
• Redistribution: derived-only

Release Contents

• U
• Y
• emotion_names
• gold_raw_scores
• instance_id
• pred_raw_scores

Provenance And Usage Notes

• Source asset: SemEval-2007 Task 14 AffectiveText
• Metadata note: The package omits raw headlines and raw API responses, and keeps only derived score arrays.

Limitations

The reported benchmark tables use a frozen closed-model cache; the release omits raw headlines and raw API responses. A fully open TF-IDF+SVD+kNN fallback cache-builder is included for local reproduction, but it is not identical to the main predictor used in the paper tables.