vi-ner-videberta
ViDeBERTa token-classification model for Vietnamese bank-statement OCR. It extracts
PERSON, ORGANIZATION, and ADDRESS; the address label is deliberately retained
to disambiguate names inside administrative addresses (for example, a person-like
street or ward name) from actual people and organizations.
Evaluation
Trainer token/window BIO metrics
These legacy Trainer summaries are computed over tokenized windows. They are reported separately from final decoded document-span exact-match metrics.
| Metric | Validation | Test |
|---|---|---|
| precision | 0.895407 | - |
| recall | 0.908239 | - |
| f1 | 0.901777 | - |
| accuracy | 0.984443 | - |
Per-entity F1
| Entity | Validation | Test |
|---|---|---|
| PERSON | 0.891723 | - |
| ORGANIZATION | 0.889339 | - |
| ADDRESS | 0.922978 | - |
The training run did not record a public dashboard URL. Local TensorBoard event files are included under tensorboard/
when present.
Training data
Split sizes: test: 29,587, train: 331,468, validation: 31,809.
| Source | Rows |
|---|---|
curated_ambiguity | 54 |
curated_statement_patterns | 10 |
llm_statement_synthetic | 1,955 |
llm_statement_wave03 | 324 |
llm_statement_wave04 | 410 |
llm_statement_wave05 | 63 |
llm_statement_wave06 | 58 |
llm_statement_wave07 | 102 |
llm_statement_wave09 | 32 |
llm_statement_wave10 | 33 |
llm_statement_wave11 | 29 |
llm_statement_wave12 | 32 |
llm_statement_wave13 | 24 |
llm_statement_wave14 | 16 |
llm_statement_wave15 | 3 |
llm_statement_wave16 | 15 |
llm_statement_wave17 | 10 |
llm_statement_wave18 | 8 |
llm_statement_wave19 | 11 |
llm_statement_wave20 | 3 |
llm_statement_wave21 | 12 |
llm_statement_wave22 | 4 |
llm_statement_wave23 | 12 |
llm_statement_wave24 | 3 |
llm_statement_wave25 | 11 |
llm_statement_wave26 | 1 |
llm_statement_wave27 | 12 |
llm_statement_wave28 | 1 |
llm_statement_wave29 | 2 |
llm_statement_wave30 | 1 |
llm_statement_wave31 | 6 |
llm_statement_wave33 | 2 |
llm_statement_wave34 | 4 |
llm_statement_wave35 | 2 |
llm_statement_wave37 | 3 |
llm_statement_wave38 | 5 |
llm_statement_wave39 | 13 |
llm_statement_wave40 | 4 |
llm_statement_wave42 | 4 |
llm_statement_wave44 | 3 |
llm_statement_wave45 | 3 |
llm_statement_wave46 | 2 |
masothue_llm_wave03 | 107 |
masothue_llm_wave04 | 72 |
masothue_statement_augmented | 34,998 |
masothue_synthetic | 119,794 |
meddies_pii_vi | 15,212 |
news_ner | 2,973 |
pap_ner | 34,103 |
phoner_covid19 | 9,955 |
private_statement_csv_llm | 7,778 |
private_statement_ocr_llm | 5,769 |
statement_hard_cases | 344 |
vietnamnet_gold | 8,396 |
vlsp2016 | 16,763 |
vlsp2021_ndtands | 19,329 |
wikiann_vi | 27,364 |
This checkpoint was trained on a mixed-source dataset that includes restricted or non-redistributable examples. The source rows are not uploaded with this model. Review every upstream license and usage condition before commercial use or redistribution.
Reproducibility
- Base checkpoint:
Fsoft-AIC/videberta-base - Base revision:
d72e04d6e5065b2babb6bccad6f86960701f0572 - Seed:
20260927 - Labels:
O,B-PERSON,I-PERSON,B-ORGANIZATION,I-ORGANIZATION,B-ADDRESS,I-ADDRESS
Intended use and limitations
The intended unit is one OCR-extracted bank-statement transaction line. OCR errors, abbreviations, unseen bank formats, and names that are also locations can reduce accuracy. Predictions should support downstream analysis, not replace review for legal, compliance, identity, or payment decisions.