v1-deberta_large-token_level-1-7

v1-deberta_large-token_level-1-7 (legacy: knowledgator/gliner-pii-large-v1.0 fine-tuned 8k steps)

Superseded. This is an early checkpoint, kept because it has a DOI (10.57967/hf/8522). For current pt-BR PII detection use arthrod/gliner-mmbert-small-ptbr-pii-full-3x-v1 or the OpenAI Privacy Filter fine-tune arthrod/gliner-opf-ptbr-pii-v1; results for both are in arthrod/gliner-opf-ptbr-pii-bench-v1.

TL;DR. The main run of the 24-Feb-2026 MI300X session documented in arthrod/attempt_vanilla (run gliner-pii-large-full-v1, config vanilla_f). It continued knowledgator/gliner-pii-large-v1.0 (DeBERTa-v3-large, token-level GLiNER) on ~1.03M in-house synthetic PII examples. Training was stopped by hand at step 8,000 of a planned 15,000 (1.44 epochs). The repo was first published as arthrod/gliner-pii-large-full-v1, and that URL still redirects here.

Among the five legacy repos in this group, this is the only one with task metrics. On the 23-label PT-BR canonical set it reaches F1 0.730, up from 0.459 for the knowledgator base. On nvidia/Nemotron-PII it reaches F1 0.627 (base 0.581). It loses ground on general and multilingual NER sets (see Evaluation).

v1, deberta_large and token_level in the repo name are accurate. 1-7 is not explained anywhere in the repo.

Related repos

Quick start

from huggingface_hub import snapshot_download
from gliner import GLiNER

# Skip the 1.76 GB ONNX file if you only need PyTorch.
path = snapshot_download(
    "arthrod/v1-deberta_large-token_level-1-7",
    allow_patterns=["gliner_config.json", "pytorch_model.bin", "tokenizer*.json"],
)
model = GLiNER.from_pretrained(path)

text = "Meu nome é Maria Silva, CPF 123.456.789-09, moro na Rua das Flores, 42, em São Paulo."
labels = ["first name", "last name", "cpf document number", "location street",
          "location building number", "location city"]
for ent in model.predict_entities(text, labels, threshold=0.5):
    print(ent["text"], "->", ent["label"], round(ent["score"], 3))

The arthrod/gliner_eval_folder (private) sweep loaded this exact repo (as arthrod/gliner-pii-large-full-v1, same weights) with GLiNER on ROCm. The snippet itself was not re-run for this card. Trained with gliner==0.2.25, transformers==5.1.0.

Files

Intermediate checkpoints live in the git history. The Trainer pushed every 500 steps ("Training in progress, step N", N = 500 … 8000). Any of them can be loaded with revision=<commit>. The lowest eval loss was at step 7,500 (commit 3a3303a1). arthrod/attempt_vanilla also publishes checkpoint-7500 as a folder.

Eval loss by step

Source: gliner_logs/gliner-pii-large-full-v1/checkpoint-8000/trainer_state.json in arthrod/attempt_vanilla (this repo has no trainer state). Eval set: about 50k examples (data/eval.json cut down to 50,000). Focal loss with sum reduction.

Best checkpoint by recorded metrics: by eval loss, step 7,500 (in git history and in attempt_vanilla). Task metrics (below) exist only for step 8,000, the weights in this repo, and for step 1,000. Step 8,000 beats step 1,000 on the three PII sets (both PT-BR sets and Nemotron-PII). It is worse on the four general-NER sets.

Training data

data/train.json: about 1,026,320 examples of in-house synthetic PII, not published. data/eval.json was cut to 50,000 examples (per attempt_vanilla). The label set is not recorded here. The strong results on the 23-label PT-BR canonical set (labels such as cpf document number, rg document number, pis document number) suggest PT-BR PII was in the mix.

Training recipe

From gliner_config.json / training_args.bin (inspected without unpickling) and attempt_vanilla's notes:

• Init knowledgator/gliner-pii-large-v1.0; encoder microsoft/deberta-v3-large (24 layers, hidden 1024); span_mode: token_level, max_len 768, max_width 100, max_types 30, max_neg_type_ratio 1, hidden_size 512, dropout 0.1, 1 RNN layer, subtoken_pooling: first, whitespace splitter.
• AdamW (torch); LR encoder 5e-6, LR others 7e-6; weight decay 0.01; linear schedule, 5% warmup; max_steps 15000 (stopped at 8,000); batch 184 (eval 176); grad clip 1.0; bf16; gradient checkpointing; seed 42.
• Focal loss alpha 0.75, gamma 0, sum reduction; negatives 2; masking: global; random_drop on.
• Eval/save every 500 steps; W&B project huggingface.
• Hardware: one AMD Instinct MI300X (192 GB), ROCm 7.1, about 6.5 h.

Evaluation

Source: data/summary.parquet of arthrod/gliner_eval_folder (private) (Feb-2026 sweep, run_eval.py). It uses GLiNER's built-in evaluator: exact span + exact label, micro precision/recall/F1. Threshold 0.5, up to 25 labels per batch (seed 42), batch 256, MI300X. Each set is sub-sampled to 5,000 examples, except ptbr_pii_val_19k (19,164).

The sweep also evaluated attempt_vanilla's local checkpoint-8000, which should be the same weights. It matches this repo exactly on 6 sets and differs slightly on Nemotron-PII (0.6155) and ptbr_pii_val_19k (0.3850). The cause is not recorded.

Per-label results on the PT-BR canonical set (same source, per_entity_json), sorted by gold count:

The pattern is high recall and low precision on structured identifiers (dob, CPF, PIS, RG, building number, middle name): the model over-predicts them. The gains are largest on the sensitive "subject described …" categories, where the base model scored near zero.

Limitations

• Trained on synthetic data only. Not a compliance tool, so do not rely on it as the sole PII safeguard.
• Over-predicts structured identifiers (precision 0.30–0.48 on PIS/CPF/dob/middle name, see above). Tune the threshold per label.
• Regresses vs. the base model on general-domain and multilingual NER sets.
• The recorded best eval-loss step (7,500) is not the one in this repo (8,000).
• The eval sets gliner-canonical-ptbr-pii-v1 and ptbr_pii_val_19k are not public.

License

Apache-2.0, matching knowledgator/gliner-pii-large-v1.0 (Apache-2.0). The encoder microsoft/deberta-v3-large is MIT.

Citation

@misc{arthrod_v1_deberta_large_token_level,
  author = {arthrod},
  title  = {v1-deberta_large-token_level-1-7 (gliner-pii-large-full-v1): knowledgator/gliner-pii-large-v1.0 fine-tuned for PII},
  year   = {2026},
  doi    = {10.57967/hf/8522},
  url    = {https://huggingface.co/arthrod/v1-deberta_large-token_level-1-7}
}

@misc{stepanov2024gliner,
  title         = {GLiNER multi-task: Generalist Lightweight Model for Various Information Extraction Tasks},
  author        = {Stepanov, Ihor and Shtopko, Mykhailo},
  year          = {2024},
  eprint        = {2406.12925},
  archivePrefix = {arXiv}
}