ELVISIO/ee5438-g36-hc-roberta-head-ablation

CAI Haochen's RoBERTa readout ablation

These three small heads rank the same frozen list of up to eight legal spans. They use the published EE5438 Group 36 RoBERTa-large encoder, the fixed three-fold out-of-fold training candidates, and training seed 58561440. The original split and encoder seeds remain unchanged.

Head features	Validation word Jaccard
Start and end vectors	0.70466
Start, end, and their interaction	0.70534
Interaction plus sentiment-marker vector	0.70338

The shared top-1 candidate scores 0.69999 and the candidate oracle scores 0.92052 on the 2,748-example validation split. The interaction-minus-boundary paired gain is 0.00068, with a 95% tweet-bootstrap interval of [-0.00673, 0.00805]. These single-seed results do not establish a reliable improvement, and no new local-test or Kaggle score is claimed.

Each joint_head.pt stores state_dict, hidden_size, and features. Load it with JointScorer from the hc-roberta branch and the exact published RoBERTa encoder; it is not a standalone text model. The candidate-list contract and executed notebooks are in the project repository. Upstream model and dataset terms continue to apply.

Project code: https://github.com/ELVISIO119/ee5438-group36-tweet-sentiment-extraction/tree/hc-roberta