Views
No views yet
google-bert/bert-base-multilingual-cased fine-tuned for Persian/Iranian PII token classification.1{
2 "model": "artifacts/google-mbert-pii-4bit/litert-int4-b16/model.int4.weight-only.tflite",
3 "dataset": "data/final_splits_audited/combined_clean",
4 "split": "test",
5 "rows": 64,
6 "max_length": 256,
7 "batch_size": 16,
8 "precision": 0.9577735124760077,
9 "recall": 0.9577735124760077,
10 "f1": 0.9577735124760077,
11 "accuracy": 0.9935635598465157
12}1{
2 "status": "converted_int4_weight_only",
3 "fp32_artifact": "artifacts/google-mbert-pii-4bit/litert-int4-b16/model.fp32.tflite",
4 "max_length": 256,
5 "batch_size": 16,
6 "fp32": {
7 "name": "litert_fp32",
8 "shape": [
9 16,
10 256,
11 39
12 ],
13 "argmax_match_rate_vs_torch": 1.0,
14 "max_abs_diff_vs_torch": 0.00022459030151367188,
15 "mean_abs_diff_vs_torch": 8.042213721637381e-07
16 },
17 "int4_artifact": "artifacts/google-mbert-pii-4bit/litert-int4-b16/model.int4.weight-only.tflite",
18 "int4_recipe": "ai_edge_quantizer.recipe.weight_only_wi4_afp32",
19 "int4": {
20 "name": "litert_int4_weight_only",
21 "shape": [
22 16,
23 256,
24 39
25 ],
26 "argmax_match_rate_vs_torch": 0.91015625,
27 "max_abs_diff_vs_torch": 9.644926071166992,
28 "mean_abs_diff_vs_torch": 0.20304147899150848
29 }
30}max_length=256 and stride around 96;کد ملی, شماره تماس, کدپستی, and ایمیل..tflite graph is fixed shape: batch 16, sequence 256. Tokenize
windows to that shape, pad the final short batch by duplicating rows, run
ai_edge_litert.Interpreter, then drop the duplicate row during postprocessing.
"""