tolyho/dl2-hw2-bge-ner

BGE NER — DL2 homework 02

Fine-tuned BAAI/bge-small-en-v1.5 for nine CoNLL-2003 BIO labels. The synthetic-label run was selected by full validation F1, without selecting on test scores.

Run	Train rows	Validation F1	Test precision	Test recall	Test F1
Baseline	10000	0.912082	0.854063	0.883853	0.868703
With synthetic labels	10010	0.916403	0.849453	0.880135	0.864522

Both comparison runs used seed 42, learning rate 5.373713206635395e-5, batch size 4 per GPU, two Tesla T4 GPUs, FP16, weight decay 0.01, and three epochs. Checkpoints were selected by validation F1. Padding and special-token labels use -100. Twenty Optuna trials used seeded subsets of 2048 training and 512 validation examples.

The ten extra labels were generated with Qwen/Qwen2.5-7B-Instruct-AWQ. These structurally valid labels include errors; synthetic augmentation did not improve test F1 in this single paired comparison.

Experiments: baseline, synthetic. The original training histories were imported into Comet after training; no models were retrained for publication.

Saved Kaggle execution (private).

from transformers import pipeline

ner = pipeline(
    "token-classification",
    model="tolyho/dl2-hw2-bge-ner",
    aggregation_strategy="simple",
)
print(ner("John works in London."))

Assignment.