wrathofgod/movie-scene-module0-v3

Module 0 Scene Encoder (v3 — Supervised Contrastive)

Joint pretraining for movie-scene understanding with:

• supervised contrastive learning with label-key positives (no heuristic noise) and full-dataset label-based negatives (no 200-sample window)
• full-batch SupCon so every in-batch same-key scene contributes as a positive
• masked language modelling on annotation scene_text
• auxiliary scene-label prediction (SmoothL1 regression, label-smoothed classification)

Backbone

• base_model: microsoft/deberta-v3-base
• embedding_dim: 256

Dataset Summary

• all_scenes: 11118
• train_scenes: 9593
• val_scenes: 1525
• all_films: 80
• train_positive_avg_candidates: 6.0552
• val_positive_avg_candidates: 3.2643
• mlm_chunks: 15118

Best Validation Metrics

• val_loss: 4.9684
• val_contrastive_loss: 4.4305
• val_aux_loss: 0.8275
• val_r_at_1: 0.0571
• val_temperature: 0.1671
• avg_cls_acc: 0.4942
• avg_reg_mae: 0.9264
• avg_bin_acc: 0.7480
• val_task/emotional_valence: 1.1910
• val_task/scene_interaction_tone: 1.3215
• val_task/conflict_nature: 1.3526
• val_task/acoustic_space: 1.5992
• val_task/reality_layer: 0.3977
• val_task/score_dynamic_shape: 1.2102
• val_task/narrative_arc_position: 1.1923
• val_task/foreshadowing_type: 1.2925
• val_task/transition_type: 1.3077
• val_task/scene_tension_raw: 0.5136
• val_task/scene_arousal: 0.4495
• val_task/scene_valence_continuous: 0.4300
• val_task/pacing_intensity: 0.4838
• val_task/action_intensity: 0.6442
• val_task/emotional_shift_trigger: 0.5622
• val_task/emotion_tags: 0.4309

Intended Use

Use this encoder as the backbone for Module 1 by setting:

CFG["backbone"] = "/kaggle/working/module0_backbone"