kunikrubika/dl2-hw2-ner

DL2 HW2 NER

BGE-small token classifier trained for named entity recognition (PER, LOC, ORG, MISC) using BIO labels.

Training data: 10,000 labeled examples from voorhs/conll2003-corrupted and 9 additional examples annotated with Qwen/Qwen2.5-7B-Instruct-AWQ. Validation and test data were unchanged. Optuna searched 20 trials on 2,048 training / 512 validation examples. Selected learning rate: 5.373713206635395e-5; per-device batch size: 4; 3 epochs; seed: 42.

Validation (3,250 examples): precision 0.9001966568, recall 0.9244362168, F1 0.9121554301, accuracy 0.9810952071. Baseline validation F1 without extra annotations: 0.9111571962. These are validation results from one run, not test-set results.

Selected by validation F1. The small observed improvement does not establish a consistent benefit across random seeds.

Training report: https://www.comet.com/kunikrubika05/dl2-hw2/0558007540974ca99754b664376ebeed