c3750-mdeberta_base-vanilla-1-2 (legacy span-level GLiNER on mDeBERTa-v3-base)

Superseded. This is an early experiment, kept because it has a DOI (10.57967/hf/8513). 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 legacy "nightmare" PII run from August 2025: a classic span-level GLiNER (span_mode: markerV0, the "vanilla" in the repo name) on microsoft/mdeberta-v3-base, saved every 250 steps up to step 3,750 (the "c3750" in the name). The run stopped at step 3,750 of a planned 15,000. The only metric recorded is eval loss, and it got worse after fine-tuning started: 4,409 at step 0, then 19.6k–21.8k at every saved checkpoint. The weights at the repo root are a copy of checkpoint-1250, not checkpoint-3750. No task metric (F1) was ever computed for any checkpoint. Treat this repo as an experiment record.

What the other name tokens mean is not recorded anywhere in the repo: 1-2 is unknown.

Related repos

Quick start

GLiNER.from_pretrained("arthrod/c3750-…") may pull the whole snapshot (about 55 GB, 15 checkpoints). Download only what you need first. GLiNER 0.2.25 has no subfolder argument, so pass a local path:

from huggingface_hub import snapshot_download
from gliner import GLiNER

repo = "arthrod/c3750-mdeberta_base-vanilla-1-2"
# Root weights (a copy of checkpoint-1250):
path = snapshot_download(repo, allow_patterns=["gliner_config.json", "pytorch_model.bin", "tokenizer*.json"])
model = GLiNER.from_pretrained(path)

# Or any saved checkpoint, e.g. the one with the lowest post-training eval loss (2750):
# path = snapshot_download(repo, allow_patterns=["checkpoint-2750/*"], ignore_patterns=["*optimizer*", "*rng_state*", "*scheduler*"])
# model = GLiNER.from_pretrained(f"{path}/checkpoint-2750")

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

This snippet was not run against these weights for this card. The config was written by transformers 4.55.0.

Files and checkpoint layout

Checkpoints

The gliner config is identical in all 15 checkpoints. Eval loss is from checkpoint-3750/trainer_state.json. "Mean train loss" is the mean of the per-step train loss over the 250 steps before the checkpoint (computed from the same file). Losses use sum reduction, so they scale with batch size and number of entity types.

Which checkpoint is best? Going only by recorded metrics: none of the fine-tuned checkpoints beats the starting point (step 0). Among the saved ones, checkpoint-2750 has the lowest eval loss. The root copy (checkpoint-1250) is mid-pack. Train loss kept falling while eval loss stayed about 20x higher. That pattern points to a train/eval mismatch or overfitting, not to a model that improved.

Training data

Not documented in the repo. What is known:

• pre_eval.log loads convert2.jsonl (50,264 items) on a ROCm box (/home/rocm-user/jupyter/…).
• trainer_state.json gives 335 optimizer steps per epoch at batch size 128. That works out to about 42.9k training examples. Eval throughput (~70 samples/s × ~67 s) puts the eval split at about 4.75k examples. Both are derived numbers, consistent with a train/eval split of a ~50k-item file.
• An earlier version of this card called it a "legacy multi-PII experiment". It pointed to a private arthrod/v1-mdeberta_base-vanilla-1-1 as "the clean v1 release".

Training recipe

From gliner_config.json and checkpoint-3750/trainer_state.json:

• Encoder microsoft/mdeberta-v3-base (12 layers, hidden 768, vocab 250,105), GLiNER span_mode: markerV0, max_width 12, hidden_size 512, dropout 0.4, RNN on (has_rnn: true), max_len 384, max_types 25, subtoken_pooling: first, whitespace word splitter.
• Initialization (inferred): every field that both configs have matches urchade/gliner_multi_pii-v1's gliner_config.json. That includes its idiosyncratic leftovers (name: "correct", num_steps: 30000, warmup_ratio: 5000, eval_every: 5000, train_batch_size: 8, lr_encoder: "5e-6"). The low step-0 eval loss also fits a pretrained GLiNER head rather than a random one. So this run most likely continued from urchade/gliner_multi_pii-v1. Those config fields are inherited metadata, not the hyper-parameters that were actually used.
• Trainer (from trainer_state.json): batch size 128, max_steps 15,000 (45 epochs), eval and save every 250 steps, logging every step. The LR ramps linearly at 3.33e-8 per step (about 1,500 warmup steps to ~5e-5). The run stopped at step 3,750 (epoch 11.19).
• Date: 2025-08-07 (initial commit and pre_eval.log).

Evaluation

There is no task-level evaluation (precision/recall/F1) for any checkpoint. The pre-evaluation in pre_eval.log never produced a score. An April-2026 version of this card showed a cross-model leaderboard in which this model's row was empty ("-"). The table is still in the repo history (commit d459fe58).

training_log_history.parquet

3,766 rows (3,750 train + 16 eval). Examples:

[
  {"kind": "train", "step": 3750, "epoch": 11.194, "loss": 1080.1953, "grad_norm": 414.35, "learning_rate": 4.665e-05, "eval_loss": null, "eval_runtime": null, "eval_samples_per_second": null},
  {"kind": "eval", "step": 3750, "epoch": 11.194, "loss": null, "grad_norm": null, "learning_rate": null, "eval_loss": 20677.736, "eval_runtime": 67.403, "eval_samples_per_second": 70.561}
]

import polars as pl
from huggingface_hub import hf_hub_download
p = hf_hub_download("arthrod/c3750-mdeberta_base-vanilla-1-2", "training_log_history.parquet")
df = pl.read_parquet(p)

Limitations

• No F1/precision/recall exists, and the only recorded metric (eval loss) got worse during fine-tuning. Do not use for real PII redaction.
• Training data, label set and language mix are not recorded. Earlier cards listed pt and en.
• The repo name says 3,750, but the root weights are checkpoint-1250.
• This is a classic span-level (markerV0) GLiNER with max_width 12. Entities longer than 12 words cannot be predicted.

License

MIT, as declared by the owner on this repo. Upstream licenses: the encoder microsoft/mdeberta-v3-base is MIT. The likely parent urchade/gliner_multi_pii-v1 is Apache-2.0, so if this run did continue from it, Apache-2.0 attribution terms also apply.

Citation

@misc{arthrod_c3750_mdeberta_base_vanilla,
  author = {arthrod},
  title  = {c3750-mdeberta_base-vanilla-1-2: legacy span-level GLiNER PII checkpoints on mDeBERTa-v3-base},
  year   = {2025},
  doi    = {10.57967/hf/8513},
  url    = {https://huggingface.co/arthrod/c3750-mdeberta_base-vanilla-1-2}
}

@inproceedings{zaratiana2024gliner,
  title     = {GLiNER: Generalist Model for Named Entity Recognition using Bidirectional Transformer},
  author    = {Zaratiana, Urchade and Tomeh, Nadi and Holat, Pierre and Charnois, Thierry},
  booktitle = {NAACL},
  year      = {2024}
}