c3750-mdeberta_base-vanilla-1-2-checkpoints
All intermediate checkpoints of the June 2025 span-level GLiNER (span_mode: markerV0) PT-BR PII run on
microsoft/mdeberta-v3-base, consolidated on 2026-10-11 from twelve single-checkpoint repos into one folder per
checkpoint. The run was planned for 37,968 steps (8 epochs); checkpoints from step 500 to 13,500 were uploaded on
2025-06-23/24. Each folder holds the 1.15 GB pytorch_model.bin, the 2.31 GB optimizer.pt, an ONNX (opset 19)
export, gliner_config.json, the tokenizer files and the original single-repo README.
Sibling run, not the DOI'd one. The published repo arthrod/c3750-mdeberta_base-vanilla-1-2 (DOI 10.57967/hf/8513) is a separate August 2025 run (15,000 planned steps, stopped at 3,750) on the same encoder and config family; its own checkpoints stay inside that repo. This repo collects the checkpoints of the earlier June 2025 run. Neither run is recommended for use today.
Related repos
| Repo | Type | Role |
|---|---|---|
| arthrod/c3750-mdeberta_base-vanilla-1-2 | model | Published (DOI) sibling run, Aug 2025 |
| arthrod/gliner-mmbert-small-ptbr-pii-full-3x-v1 | model | Current recommended PT-BR PII GLiNER |
| arthrod/gliner-opf-ptbr-pii-bench-v1 | dataset | Benchmark of the current suite |
Checkpoints
| folder | step | files | size | former repo | trainer_state.json |
|---|---|---|---|---|---|
c500-mdeberta_base-vanilla-1-2/ |
500 | 13 | 4.65 GB | arthrod/c500-mdeberta_base-vanilla-1-2 |
yes |
gliner-mdeberta-base-ptbr-pii-run2506-ckpt1000/ |
1,000 | 13 | 4.65 GB | arthrod/gliner-mdeberta-base-ptbr-pii-run2506-ckpt1000 |
yes |
gliner-mdeberta-base-ptbr-pii-run2506-ckpt1500/ |
1,500 | 13 | 4.65 GB | arthrod/gliner-mdeberta-base-ptbr-pii-run2506-ckpt1500 |
yes |
gliner-mdeberta-base-ptbr-pii-run2506-ckpt2000/ |
2,000 | 13 | 4.65 GB | arthrod/gliner-mdeberta-base-ptbr-pii-run2506-ckpt2000 |
yes |
gliner-mdeberta-base-ptbr-pii-run2506-ckpt2500/ |
2,500 | 13 | 4.65 GB | arthrod/gliner-mdeberta-base-ptbr-pii-run2506-ckpt2500 |
yes |
c3k-mdeberta_base-vanilla-1-2/ |
3,000 | 13 | 4.65 GB | arthrod/c3k-mdeberta_base-vanilla-1-2 |
yes |
c3500-mdeberta_base-vanilla-1-2/ |
3,500 | 13 | 4.65 GB | arthrod/c3500-mdeberta_base-vanilla-1-2 |
yes |
c4k-mdeberta_base-vanilla-1-2/ |
4,000 | 13 | 4.65 GB | arthrod/c4k-mdeberta_base-vanilla-1-2 |
yes |
c4500-mdeberta_base-vanilla-1-2/ |
4,500 | 13 | 4.65 GB | arthrod/c4500-mdeberta_base-vanilla-1-2 |
yes |
c5k-mdeberta_base-vanilla-1-2/ |
5,000 | 13 | 4.65 GB | arthrod/c5k-mdeberta_base-vanilla-1-2 (renamed into this repo) |
yes |
c5500-mdeberta_base-vanilla-1-2/ |
5,500 | 13 | 4.65 GB | arthrod/c5500-mdeberta_base-vanilla-1-2 |
yes |
gliner-mdeberta-base-ptbr-pii-run2506-ckpt13500/ |
13,500 | 13 | 4.65 GB | arthrod/gliner-mdeberta-base-ptbr-pii-run2506-ckpt13500 |
yes |
Every file was copied byte-for-byte (sha256 verified against the source repo before the source was deleted).
The old repo names do not redirect, except c5k-mdeberta_base-vanilla-1-2, which was renamed into this repo.
How to use
Download one folder only (the whole repo is about 56 GB):
from huggingface_hub import snapshot_download
from gliner import GLiNER
path = snapshot_download("arthrod/c3750-mdeberta_base-vanilla-1-2-checkpoints", allow_patterns=["c3k-mdeberta_base-vanilla-1-2/*"],
ignore_patterns=["*/optimizer.pt"])
model = GLiNER.from_pretrained(f"{path}/c3k-mdeberta_base-vanilla-1-2")
model.predict_entities("Meu CPF é 123.456.789-09 e moro em Curitiba.", ["cpf", "cidade"], threshold=0.5)
optimizer.pt is only needed to resume training.
Limitations
- Mid-run checkpoints of an experiment; no task metric (F1) was computed for any of them. Each folder's
trainer_state.jsonholds the loss log up to that step (the logs are prefixes of one another). 1-2in the name is a legacy run tag whose meaning is not recorded.- Superseded by the 2026 PT-BR PII models linked above.