Token Classification
Transformers
Safetensors
English
bert
ner
THemidli commited on
Commit
bd52f45
·
verified ·
1 Parent(s): cfd5dfe

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +69 -0
README.md ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: google/bert_uncased_L-2_H-128_A-2
3
+ datasets:
4
+ - THemidli/applied-ner-stage2-expanded
5
+ language:
6
+ - en
7
+ library_name: transformers
8
+ pipeline_tag: token-classification
9
+ license: apache-2.0
10
+ tags:
11
+ - ner
12
+ - token-classification
13
+ ---
14
+
15
+ # Applied NER Stage 3 — BERT Tiny
16
+
17
+ An eight-label English token classifier fine-tuned from [google/bert_uncased_L-2_H-128_A-2](https://huggingface.co/google/bert_uncased_L-2_H-128_A-2). Repository: THemidli/applied-ner-stage3-bert-tiny.
18
+
19
+ ## Results
20
+
21
+ Exact entity-level seqeval metrics:
22
+
23
+ | Split | Precision | Recall | F1 | Token accuracy |
24
+ |---|---:|---:|---:|---:|
25
+ | Train | 0.9581 | 0.9726 | 0.9653 | 0.9957 |
26
+ | Test | 0.4215 | 0.5273 | 0.4685 | 0.8269 |
27
+
28
+ | Label | Precision | Recall | F1 | Support |
29
+ |---|---:|---:|---:|---:|
30
+ | PERSON | 0.461 | 0.641 | 0.536 | 195 |
31
+ | ORGANIZATION | 0.200 | 0.218 | 0.208 | 147 |
32
+ | LOCATION | 0.432 | 0.552 | 0.485 | 143 |
33
+ | TIMEDATE | 0.792 | 0.844 | 0.817 | 167 |
34
+ | PRODUCT | 0.171 | 0.189 | 0.180 | 127 |
35
+ | WORKOFART | 0.128 | 0.247 | 0.168 | 97 |
36
+ | JOB | 0.699 | 0.798 | 0.745 | 99 |
37
+ | AMOUNT | 0.556 | 0.625 | 0.588 | 104 |
38
+
39
+ On 40 fresh, manually gold-labeled wild probes, exact span F1 was 0.4785 (precision 0.4505, recall 0.5102).
40
+
41
+ ## Training
42
+
43
+ - Dataset: [THemidli/applied-ner-stage2-expanded](https://huggingface.co/datasets/THemidli/applied-ner-stage2-expanded)
44
+ - Seed: 20260802
45
+ - Hardware: Apple MPS (macOS-27.0-arm64-arm-64bit)
46
+ - Runtime: 17.780 seconds
47
+ - Records/chunks: 641/665 train; 159/165 test
48
+ - Maximum length: 256; fast-tokenizer overflow chunks, no overlapping stride
49
+ - Hyperparameters: {"epochs": 20, "eval_batch_size": 64, "learning_rate": 0.0005, "scheduler": "linear", "train_batch_size": 32, "warmup_steps": 42, "weight_decay": 0.01}
50
+ - No validation split and no test-driven checkpoint selection
51
+
52
+ ## Footprint and CPU benchmark
53
+
54
+ - Parameters: 4,371,601 (17.49 MB tensor storage)
55
+ - Saved artifact: 18.21 MB
56
+ - Model-load RSS delta: 31.85 MB
57
+ - End-to-end inference RSS delta: 41.30 MB
58
+ - CPU throughput: 11459.9 examples/s at batch 32 with 8 threads
59
+ - Mean latency: 0.0873 ms/example at that batch size
60
+
61
+ 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.
62
+
63
+ ## Labels
64
+
65
+ PERSON, ORGANIZATION, LOCATION, TIMEDATE, PRODUCT, WORKOFART, JOB, AMOUNT using BIO encoding.
66
+
67
+ ## Limitations
68
+
69
+ 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.