Applied NER Stage 4 — Final BERT-Mini

An eight-label English token classifier fine-tuned from prajjwal1/bert-mini. Repository: THemidli/applied-ner-stage4-bert-mini-final.

Results

Exact entity-level seqeval metrics:

On 40 fresh, manually gold-labeled wild probes, exact span F1 was 0.7685 (precision 0.7429, recall 0.7959). Test F1 changed by +0.1771 versus Stage 3.

On the MPS backend, repeated runs with identical seed and config showed ±0.01–0.015 F1 variation (0.6331 vs 0.6455 across the two recorded runs); the seed is fixed, and the variation does not change the model ranking (both runs far above BERT-Tiny, slightly below ELECTRA-Small).

Training

• Dataset: THemidli/applied-ner-stage4-final
• Seed: 20260802
• Hardware: Apple MPS (macOS-27.0-arm64-arm-64bit)
• Runtime: 32.908 seconds
• Records/chunks: 841/865 train; 159/165 test
• Maximum length: 256; fast-tokenizer overflow chunks, no overlapping stride
• Hyperparameters: {"attention_dropout": 0.1, "classifier_dropout": 0.1, "epochs": 16, "eval_batch_size": 64, "hidden_dropout": 0.1, "label_smoothing_factor": 0.0, "learning_rate": 0.0005, "scheduler": "linear", "train_batch_size": 32, "warmup_steps": 45, "weight_decay": 0.02}
• No validation split and no test-driven checkpoint selection

Footprint and CPU benchmark

• Parameters: 11,109,137 (44.44 MB tensor storage)
• Saved artifact: 45.16 MB
• Model-load RSS delta: 32.29 MB
• End-to-end inference RSS delta: 58.52 MB
• CPU throughput: 4141.6 examples/s at batch 32 with 8 threads
• Mean latency: 0.2415 ms/example at that batch size

The benchmark covers tokenizer plus PyTorch CPU forward pass over 40 short probes, repeated 50 times. It is workload- and hardware-specific, not single-request latency.

Labels

PERSON, ORGANIZATION, LOCATION, TIMEDATE, PRODUCT, WORKOFART, JOB, AMOUNT using BIO encoding.

Limitations

This is an 11.11M-parameter uncased four-layer BERT trained on a small, heterogeneous dataset with controlled template additions. It is not a production privacy system. PRODUCT, WORKOFART, ORGANIZATION, title boundaries, and contextual site readings remain weak. The 40-probe wild set is diagnostic, not a population benchmark.