Applied NER Stage 3 — BERT Tiny
An eight-label English token classifier fine-tuned from google/bert_uncased_L-2_H-128_A-2. Repository: THemidli/applied-ner-stage3-bert-tiny.
Results
Exact entity-level seqeval metrics:
| Split | Precision | Recall | F1 | Token accuracy |
|---|---|---|---|---|
| Train | 0.9581 | 0.9726 | 0.9653 | 0.9957 |
| Test | 0.4215 | 0.5273 | 0.4685 | 0.8269 |
| Label | Precision | Recall | F1 | Support |
|---|---|---|---|---|
| PERSON | 0.461 | 0.641 | 0.536 | 195 |
| ORGANIZATION | 0.200 | 0.218 | 0.208 | 147 |
| LOCATION | 0.432 | 0.552 | 0.485 | 143 |
| TIMEDATE | 0.792 | 0.844 | 0.817 | 167 |
| PRODUCT | 0.171 | 0.189 | 0.180 | 127 |
| WORKOFART | 0.128 | 0.247 | 0.168 | 97 |
| JOB | 0.699 | 0.798 | 0.745 | 99 |
| AMOUNT | 0.556 | 0.625 | 0.588 | 104 |
On 40 fresh, manually gold-labeled wild probes, exact span F1 was 0.4785 (precision 0.4505, recall 0.5102).
Training
- Dataset: THemidli/applied-ner-stage2-expanded
- Seed: 20260802
- Hardware: Apple MPS (macOS-27.0-arm64-arm-64bit)
- Runtime: 17.780 seconds
- Records/chunks: 641/665 train; 159/165 test
- Maximum length: 256; fast-tokenizer overflow chunks, no overlapping stride
- Hyperparameters: {"epochs": 20, "eval_batch_size": 64, "learning_rate": 0.0005, "scheduler": "linear", "train_batch_size": 32, "warmup_steps": 42, "weight_decay": 0.01}
- No validation split and no test-driven checkpoint selection
Footprint and CPU benchmark
- Parameters: 4,371,601 (17.49 MB tensor storage)
- Saved artifact: 18.21 MB
- Model-load RSS delta: 31.85 MB
- End-to-end inference RSS delta: 41.30 MB
- CPU throughput: 11459.9 examples/s at batch 32 with 8 threads
- Mean latency: 0.0873 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 a 4.37M-parameter uncased two-layer BERT trained on a small, heterogeneous dataset. It is a compact baseline, not a production privacy system. Rare works/products, company-versus-product context, exact boundaries, and subword-heavy names remain weak. The 40-probe wild set is diagnostic, not a population benchmark.