movie-scene-module0

Module 0 Scene Encoder

Joint pretraining for movie-scene understanding with:

• supervised contrastive learning
• masked language modeling on annotation scene_text
• auxiliary scene-label prediction

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: 8.0
• val_positive_avg_candidates: 7.9993
• mlm_chunks: 13596

Best Validation Metrics

• val_loss: 3.0135
• val_contrastive_loss: 2.5612
• val_aux_loss: 0.6957
• val_r_at_1: 0.1348
• avg_cls_acc: 0.5545
• avg_reg_mae: 0.8708
• avg_bin_acc: 0.6279
• val_task/emotional_valence: 1.0642
• val_task/scene_interaction_tone: 1.1227
• val_task/conflict_nature: 1.1504
• val_task/acoustic_space: 1.4399
• val_task/reality_layer: 0.2565
• val_task/score_dynamic_shape: 1.2453
• val_task/narrative_arc_position: 1.0132
• val_task/foreshadowing_type: 1.2081
• val_task/transition_type: 1.1950
• val_task/scene_tension_raw: 0.3548
• val_task/scene_arousal: 0.3117
• val_task/scene_valence_continuous: 0.2986
• val_task/pacing_intensity: 0.2708
• val_task/action_intensity: 0.3631
• val_task/emotional_shift_trigger: 0.6338
• val_task/emotion_tags: 0.4502

Intended Use

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

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