gliner-ettin-32m-ptbr-pii-full-3x-v1-checkpoints
Every earlier GLiNER run on jhu-clsp/ettin-encoder-32m for Brazilian-Portuguese PII detection that led to the published model arthrod/gliner-ettin-32m-ptbr-pii-full-3x-v1 (DOI 10.57967/hf/8519), consolidated on 2026-10-11 from four repos into one folder per run. Use the published model for inference; this repo is for reproducing the data-size ablation (top-50k vs top-100k vs full train set) and for resuming or inspecting training.
Related repos
| Repo | Type | Role |
|---|---|---|
| arthrod/gliner-ettin-32m-ptbr-pii-full-3x-v1 | model | Published model (3x budget, 69,000 steps); has the evaluation |
| arthrod/gliner-ettin-68m-ptbr-pii-full-3x-v1 | model | Published 68M sibling |
| arthrod/gliner-ettin-68m-ptbr-pii-full-3x-v1-checkpoints | model | Same consolidation for the 68M encoder |
| arthrod/gliner-mmbert-small-ptbr-pii-full-3x-v1 | model | Strongest GLiNER of the family |
| arthrod/gliner-opf-ptbr-pii-bench-v1 | dataset | Benchmark of the current suite |
Runs
| folder | run | checkpoint-N/ folders (steps) |
size |
|---|---|---|---|
gliner-ettin-32m-ptbr-pii-top50k-v1/ |
Sweep: top-50k curated train set | 21 (500–10,400) | 8.93 GB |
gliner-ettin-32m-ptbr-pii-top100k-v1/ |
Sweep: top-100k curated train set | 20 (260–5,200) | 8.51 GB |
gliner-ettin-32m-ptbr-pii-full-v1/ |
Sweep: full ~984k train set (1x budget, 23,000 steps) | 20 (1,150–23,000) | 8.51 GB |
c22k-ettin32m-token_level-0-2/ |
Earlier token-level run, 22 checkpoints (final = v1-ettin32m-token_level-0-2) |
22 (1,000–22,000) | 4.06 GB |
Each folder keeps the original repo's README (with that run's recipe and metrics) next to its files. Every file
was copied byte-for-byte; sha256 was verified against the source repo before the source was deleted. The old
names of the three copied repos no longer resolve; gliner-ettin-32m-ptbr-pii-top50k-v1 was renamed into this repo
and still redirects here.
How to use
The whole repo is about 30 GB; download one folder:
from huggingface_hub import snapshot_download
from gliner import GLiNER
ckpt = "gliner-ettin-32m-ptbr-pii-top100k-v1/checkpoint-5200" # last step of that run
path = snapshot_download("arthrod/gliner-ettin-32m-ptbr-pii-full-3x-v1-checkpoints",
allow_patterns=[f"{ckpt}/*"], ignore_patterns=["*.pt", "*.pth"])
model = GLiNER.from_pretrained(f"{path}/{ckpt}")
model.predict_entities("Meu CPF é 123.456.789-09.", ["cpf document number"], threshold=0.5)
Each run folder holds only checkpoint-N/ subfolders (plus its README). optimizer.pt, scheduler.pt and rng_state.pth
are only needed to resume training.
Limitations
- Ablation and intermediate runs: they score below the published 3x model (see each folder's README and the published card for metrics).
- Trained on synthetic PT-BR PII; all identities in the training data are generated.
gliner==0.2.25has nosubfolderargument, hence the local-path loading above.