| { |
| "model_type": "contextualized_lemmatizer", |
| "architecture": "ContextualizedLemmatizer", |
| "framework": "pytorch", |
| "language": "ang", |
| "task": "lemmatization", |
| "description": "Character-level encoder-decoder Transformer for Old English lemmatisation with a joint UPOS classification head. This is the Nerthus resource-enhanced retraining (condition C15_canonical): the architecture is unchanged from the published baseline, but the model is trained on a repaired gold standard (universe 'revised7') and decoded with a 10-best beam re-ranked by a soft prior over the released headword inventory. At inference the model is run in single-word (context-free) mode.", |
| "char_vocab_size": 140, |
| "num_pos_tags": 18, |
| "d_model": 512, |
| "nhead": 8, |
| "num_encoder_layers": 4, |
| "num_decoder_layers": 4, |
| "num_sentence_layers": 2, |
| "dim_feedforward": 1024, |
| "dropout": 0.15, |
| "max_word_length": 50, |
| "max_sentence_length": 100, |
| "num_parameters": 26514078, |
| "num_trainable_parameters": 26437278, |
| "weights_file": "model.pt", |
| "weights_format": "raw_state_dict", |
| "vocab_file": "char_vocab.json", |
| "special_tokens": { |
| "pad": "<PAD>", |
| "unk": "<UNK>", |
| "sos": "<SOS>", |
| "eos": "<EOS>" |
| }, |
| "special_token_ids": { |
| "pad": 0, |
| "unk": 1, |
| "sos": 2, |
| "eos": 3 |
| }, |
| "pos_tags": [ |
| "PAD", "NOUN", "VERB", "ADJ", "ADV", "PROPN", "PRON", "DET", |
| "ADP", "AUX", "CCONJ", "SCONJ", "NUM", "PART", "INTJ", "PUNCT", "X", "SYM" |
| ], |
| "condition": "C15_canonical", |
| "universe": "revised7", |
| "decoding": { |
| "strategy": "soft_dict_4.0", |
| "beam_size": 10, |
| "max_decode_length": 32, |
| "length_penalty": 0.0, |
| "lexicon_bonus": 4.0, |
| "inventory_file": "lemma_inventory.json", |
| "inventory_size": 42531 |
| }, |
| "input_normalisation": { |
| "lowercase": true, |
| "strip_macrons": true, |
| "eth_to_thorn": false, |
| "note": "Input word forms are lowercased and stripped of macrons before encoding. Eth is left as written in the input; it is unified with thorn in the output lemma only, because that is the convention of the training targets." |
| }, |
| "lemma_notation": { |
| "case": "lowercase", |
| "macrons": "removed", |
| "eth": "unified with thorn", |
| "form": "bare headword", |
| "note": "Lemmas are emitted in the harmonised 'revised7' convention, not in the original corpus notation. Accuracy figures in this notation are not directly comparable with the published 67.58%." |
| }, |
| "training": { |
| "epochs": 80, |
| "epochs_run": 80, |
| "best_epoch": 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", |
| "lr_range": [1e-05, 0.00015], |
| "num_cycles": 4, |
| "warmup_fraction": 0.05, |
| "early_stopping_patience": 20, |
| "label_smoothing": 0.15, |
| "pos_loss_weight": 0.2, |
| "character_data_augmentation": "attested", |
| "augmentation_probability": 0.2, |
| "pretrained_encoder": true, |
| "random_seed": 42, |
| "steps_per_epoch": 1168, |
| "total_steps": 93440, |
| "training_instances": 84050, |
| "training_condition": "corpus-only (YCOE prose and poetry plus ParCorOE), gold standard repaired in seven passes", |
| "device": "mps", |
| "wall_seconds": 9139.5 |
| }, |
| "performance": { |
| "decoding": "beam10 + soft_dict_4.0", |
| "test_lemma_accuracy": 0.8157, |
| "test_pos_accuracy": 0.8279, |
| "dev_lemma_accuracy": 0.818, |
| "dev_pos_accuracy": 0.8239, |
| "test_set_size": 10507, |
| "dev_set_size": 10506, |
| "greedy_test_lemma_accuracy": 0.7832, |
| "beam10_test_lemma_accuracy": 0.7851, |
| "oracle_in_beam10_test": 0.9537, |
| "published_baseline_test_lemma_accuracy": 0.6758, |
| "note": "Lemma accuracy is exact string match against the repaired 'revised7' targets and is not directly comparable with the published baseline's 67.58%, which is measured against the original corpus notation." |
| } |
| } |
|
|
