| { |
| "model": { |
| "architecture": "ContextualizedLemmatizer (character-level Transformer encoder-decoder with joint UPOS head; run context-free at inference)", |
| "d_model": 512, |
| "nhead": 8, |
| "num_encoder_layers": 4, |
| "num_decoder_layers": 4, |
| "num_sentence_layers": 2, |
| "dim_feedforward": 1024, |
| "dropout": 0.15, |
| "positional_encoding": "sinusoidal", |
| "char_vocab_size": 140, |
| "num_pos_tags": 18, |
| "max_word_length": 50, |
| "num_parameters": 26514078 |
| }, |
| "training": { |
| "condition": "corpus-only (YCOE)", |
| "training_instances": 84050, |
| "dev_instances": 10506, |
| "test_instances": 10507, |
| "split": "80/10/10 random, seed 42", |
| "epochs": 80, |
| "batch_size": 24, |
| "gradient_accumulation_steps": 3, |
| "effective_batch_size": 72, |
| "optimizer": "AdamW", |
| "weight_decay": 0.01, |
| "learning_rate": 0.00015, |
| "lr_schedule": "cosine annealing with warm restarts (SGDR)", |
| "lr_range": [1e-05, 0.00015], |
| "num_cycles": 4, |
| "cycle_peak_epochs": [22, 41, 58, 74], |
| "warmup_fraction": 0.05, |
| "label_smoothing": 0.15, |
| "pos_loss_weight": 0.2, |
| "early_stopping_patience": 20, |
| "best_epoch": 74 |
| }, |
| "data_augmentation": { |
| "type": "character-level, Old English spelling variation", |
| "operations": [ |
| "substitution (þ<->ð, æ<->a/e, i<->y)", |
| "deletion (words > 2 chars)", |
| "adjacent-character swap", |
| "duplication" |
| ], |
| "probability_per_word": 0.2, |
| "reapplied_each_epoch": true |
| }, |
| "performance": { |
| "dev_lemma_accuracy": 0.6868, |
| "test_lemma_accuracy": 0.6758, |
| "test_pos_accuracy": 0.8275, |
| "correct_lemmas": 7101, |
| "correct_pos": 8695, |
| "total_test_samples": 10507 |
| } |
| } |
|
|
