Ali-0111/applied-nlp-ner-stage4-distilbert

Applied NLP / NER — Stage 4: Diagnose, Relabel, Retrain

Retrained distilbert-base-cased on Ali-0111/applied-nlp-ner-stage4-retrained, same hyperparameters as Stage 3 (6 epochs, lr 5e-5, batch 16, seed 42) so the metric change is attributable to the data change alone.

Metrics comparison (Stage 3 → Stage 4, test split)

Stage 3 overall: F1=0.803 Stage 4 overall: F1=0.742

Stage 4 test set (full)

Label	Precision	Recall	F1	Support
AMOUNT	0.850	0.884	0.867	199
JOB	0.771	0.791	0.781	115
LOCATION	0.676	0.817	0.740	115
ORGANIZATION	0.634	0.717	0.673	152
PERSON	0.778	0.831	0.803	118
PRODUCT	0.240	0.600	0.343	40
TIMEDATE	0.728	0.782	0.754	188
WORKOFART	0.671	0.797	0.729	59

Overall: P=0.693, R=0.797, F1=0.742, Accuracy=0.903

Stage 4 train set (full)

Overall: P=0.933, R=0.945, F1=0.939, Accuracy=0.988

Reevaluation on the 10 Stage 3 problem patterns

See stage4_retrain/analysis/reeval.md in the GitHub repo for the full before/after comparison with concrete examples for every pattern (improved, regressed, or unresolved).