Fine-tune of openai/privacy-filter on Brazilian-Portuguese PII, 9-round chunked schedule (3 epochs × 3 saves per epoch). Trained on 914,452 rows of natural-text upstream data (arthrod/oai-pf-ptbr-chunked-v2), evaluated on the same 5,000 PT-BR val set used for the GLiNER models.

Best checkpoint: e3_c3 (final pass) — detection.span typed F1 0.885 (P 0.894 / R 0.876).

Headline performance

opf eval --eval-mode typed on the 5,000-row natural val:

detection.span typed F1 = 0.885 (P=0.894 R=0.876)

Apples-to-apples vs the GLiNER series (same val, same 24 PT-BR labels, nervaluate)

Model	partial P	partial R	partial F1	exact F1
gliner-opf-ptbr-pii-v1 (this)	0.917	0.879	0.897	0.853
mmBERT-small × 3 (41400)	0.951	0.832	0.888	0.870
ettin-68m easter-egg	0.893	0.761	0.822	0.789
ettin-32m easter-egg	0.905	0.729	0.808	0.769

This model wins partial F1 by +0.01; mmBERT wins exact F1 by +0.017. Roughly tied overall, with different strengths:

opf wins on free-text sensitive descriptors (medical +0.16, organizational +0.34, political +0.24, sexual +0.14, religious +0.04, ethnicity +0.03) mmBERT wins on structured PII + names (first/middle/last names by 0.04–0.16, locations by 0.05–0.14, full address) Both perfect on cpf/rg/pis/credit_card/phone/email/zip (F1 ≥ 0.99)

Learning curve (9-round chunked schedule)

ckpt	P	R	F1
baseline (untyped)	0.633	0.466	0.537
e1_c1	0.746	0.749	0.748
e1_c2	0.831	0.808	0.819
e1_c3 (epoch 1)	0.852	0.836	0.844
e2_c1	0.876	0.835	0.855
e2_c2	0.889	0.844	0.866
e2_c3 (epoch 2)	0.895	0.851	0.872
e3_c1	0.896	0.860	0.878
e3_c2	0.880	0.880	0.880
e3_c3 (final, released)	0.894	0.876	0.885

Per-entity F1 (e3_c3, span-typed, top entities)

label	P	R	F1
cpf_document_number	0.998	1.000	0.999
rg_document_number	0.999	0.999	0.999
phone_number	0.997	0.999	0.998
pis_document_number	0.999	0.996	0.997
dob	0.995	0.999	0.997
email_address	0.996	0.996	0.996
location_zip	0.990	1.000	0.995
credit_card	1.000	0.990	0.995
location_building_number	0.967	0.972	0.969
last_name	0.949	0.965	0.957
location_street	0.926	0.954	0.940
location_state_abbreviation	0.921	0.861	0.890
first_name	0.881	0.874	0.878
personal_description_of_ethnicity	0.847	0.842	0.844
personal_description_of_religious_convictions	0.846	0.805	0.825
personal_description_of_sexual_information	0.853	0.788	0.819
personal_description_of_political_opinion	0.821	0.815	0.818
personal_description_of_organizational_affiliation	0.800	0.794	0.797
personal_description_of_medical_conditions	0.816	0.755	0.784
location_state	0.775	0.763	0.769
location_neighborhood	0.792	0.636	0.705
location_city	0.666	0.591	0.626
middle_name	0.600	0.530	0.563

(Entries with zero gold in val are omitted.)

Training recipe

Backbone: openai/privacy-filter (8-layer MoE transformer, 128 experts, ~2.7B-equivalent params via top-4 routing)
Schedule: 3 epochs × 3 saves per epoch (9 sequential opf train --epochs 1 invocations, each on a deterministic 1/3 chunk, resuming from the previous checkpoint)
Optimizer: AdamW, LR 1e-5, weight decay 0.01, max grad norm 1.0
Batch: 32 windows × 4 grad-accum = effective 128
Context: n-ctx 256
Precision: bf16 weights, fp32 accumulators
Loss: standard CE on BIESO token labels (1 + 72 entities × 4 = 289 token labels)
Decoding: constrained Viterbi
Hardware: AMD MI300X single-GPU partition, ROCm 7.2

Dataset

Train: arthrod/oai-pf-ptbr-chunked-v2 (private) — 914,452 rows, 100% upstream raw text
- 99.8% from ai4privacy/open-pii-masking-500k-ai4privacy
- 100% from ai4privacy/pii-masking-400k
- 84.6% from arthrod/gliner2-pii-ptbr-reward-split
- 93.4% from nvidia/Nemotron-PII
- 4 small spam/phishing sources at 100% (negative evidence)
- 3 sources dropped entirely (schema mismatches, ~18.6k rows)
Val: same 5,000 PT-BR rows used for the GLiNER models — direct head-to-head comparison

Usage

python

import opf
# CLI:
# opf redact --checkpoint <download_dir> "text com cpf 123.456.789-09 e telefone (11) 91234-5678"

GLiNER series (same val): mmBERT-small × 3 (partial F1 0.823), ettin-68m-easter-egg (0.682), ettin-32m-easter-egg (0.603)
Demo: arthrod/gliner-ptbr-pii-demo
Note: easter-egg label berco-de-tiradentes is NOT supported here — use the mmBERT-small model for that.

arthrod/gliner-opf-ptbr-pii-v1

gliner-opf-ptbr-pii-v1