Create README.md

README.md

@@ -0,0 +1,116 @@
+---
+license: mit
+language:
+  - en
+library_name: pytorch
+pipeline_tag: text-classification
+base_model:
+  - roberta-large
+tags:
+  - semeval
+  - semeval2026
+  - affective-computing
+  - emotion-regression
+  - valence-arousal
+  - temporal-modeling
+datasets:
+  - semeval2026-task2
+metrics:
+  - pearsonr
+  - r_within
+  - r_between
+model-index:
+  - name: AffectDynamics-SemEval2026Task2
+    results:
+      - task:
+          type: text-classification
+          name: SemEval-2026 Task 2 (Composite)
+        dataset:
+          type: semeval2026-task2
+          name: SemEval-2026 Task 2 Validation Split
+          split: validation
+        metrics:
+          - type: r_composite
+            name: Composite Correlation
+            value: 0.6990
+---
+
+# AffectDynamics-SemEval2026Task2
+
+This repository contains a correlation-optimized temporal affect model for **SemEval-2026 Task 2**: predicting **valence** and **arousal** dynamics from user-authored essays and feeling-word entries.
+
+[![Python 3.10+](https://img.shields.io/badge/python-3.10+-blue.svg)](https://www.python.org/downloads/)
+[![PyTorch](https://img.shields.io/badge/PyTorch-2.0+-red.svg)](https://pytorch.org/)
+[![Lightning](https://img.shields.io/badge/Lightning-2.0+-purple.svg)](https://lightning.ai/)
+[![License](https://img.shields.io/badge/license-MIT-green.svg)](LICENSE)
+
+## Model Card
+
+### Model details
+
+- **Model type**: Multi-task temporal regression (Subtask 1, 2A, 2B)
+- **Backbone**: `roberta-large`
+- **Temporal encoder**: 2-layer unidirectional GRU (hidden size 384)
+- **Personalization**: Gated user embedding (24-dim)
+- **Training objective**: Correlation-first, variance-aware losses aligned with task metrics
+- **Primary checkpoint**: `best-epoch=14-val_r_composite_avg=0.6990.ckpt`
+
+### Intended use
+
+- Research use for longitudinal affect forecasting on SemEval-style data.
+- Produces continuous predictions for:
+  - Subtask 1: `pred_valence`, `pred_arousal`
+  - Subtask 2A: `pred_state_change_valence`, `pred_state_change_arousal`
+  - Subtask 2B: `pred_dispo_change_valence`, `pred_dispo_change_arousal`
+
+### Out-of-scope use
+
+- Clinical diagnosis or mental health decision support.
+- High-stakes individual-level decision making.
+- Use on domains, languages, or demographics not represented in SemEval Task 2 data without re-validation.
+
+### Training and evaluation data
+
+- Source task: SemEval-2026 Task 2 (shared-task format).
+- Training corpus in this repo includes:
+  - `data/train_subtask1.csv`
+  - `data/train_subtask2a.csv` (or computed from Subtask 1 timeline)
+  - `data/train_subtask2b_user_disposition_change.csv`
+- Validation strategy: temporal per-user split to prevent future leakage.
+
+### Metrics
+
+- **Subtask 1**: `r_within`, `r_between`, `r_composite` (per SemEval evaluator)
+- **Subtask 2A/2B**: Pearson correlation (`r`) on forecasting targets
+- **Checkpoint selection signal**: `val_r_composite_avg`
+
+### Quick start
+
+Use the provided script to download the checkpoint and generate submission files:
+
+```bash
+python generate_submission.py
+```
+
+Or run local inference with custom inputs:
+
+```bash
+python predict.py
+```
+
+### Limitations and bias
+
+- Performance depends on temporal history quality and per-user data sparsity.
+- Arousal typically has lower correlation than valence due to lower target variance.
+- Predictions are correlation-optimized for benchmark metrics and may require calibration for deployment settings.
+
+### Citation
+
+If you use this model, please cite the SemEval-2026 Task 2 shared task and this repository.
+
+## 🎯 Task Overview
+
+**Three interconnected subtasks:**
+- **Subtask 1**: Longitudinal Affect Assessment - Predict valence/arousal for each text in a user's timeline
+- **Subtask 2A**: State Change Detection - Predict short-term emotional shifts between consecutive texts  
+- **Subtask 2B**: Dispositional Change - Predict long-term changes in baseline emotional state