v1-deberta_small-token_level-0-6 (legacy, failed run: do not use)

This model does not work. The training run behind it logged a train loss, eval loss and grad-norm of exactly 0.0 from step 10 to step 17,000. The only recorded benchmark scored it 0.0000 on every metric. It is kept for the record. For pt-BR PII detection use arthrod/gliner-mmbert-small-ptbr-pii-full-3x-v1 or arthrod/gliner-opf-ptbr-pii-v1.

TL;DR. This is run h-gliner-token-full-v1 (config configs/vanilla_h.yaml in arthrod/attempt_vanilla). It tried to train a token-level GLiNER from scratch on microsoft/deberta-v3-small (prev_path: "none"), on the same ~1.03M-example synthetic PII corpus as v1-deberta_large-token_level-1-7. It ran concurrently on one MI300X on 24–25 Feb 2026. The Trainer pushed every 500 steps up to step 17,000 of a planned 60,000, and then a DataLoader worker crash ended the run. Because the loss was identically zero, the weights were never meaningfully updated by a working objective. The root cause was never confirmed. attempt_vanilla names two plausible suspects: prev_path: "none" passed as a string, and a hard-coded class_token_index.

The repo was first published as arthrod/h-gliner-token-full-v1, and that URL still redirects here. v1, deberta_small and token_level in the repo name are accurate. 0-6 is not explained anywhere in the repo.

Related repos

Loading (for forensics only)

from huggingface_hub import snapshot_download
from gliner import GLiNER

path = snapshot_download(
    "arthrod/v1-deberta_small-token_level-0-6",
    allow_patterns=["gliner_config.json", "pytorch_model.bin", "tokenizer*.json"],
)
model = GLiNER.from_pretrained(path)  # expect empty or meaningless predictions

Loading history: the arthrod/gliner_eval_folder (private) triage (late Feb 2026) failed with "Saved embeddings (128100) vs model expects (128004)". gliner_config.json was then fixed on 2026-03-19 (commit "Fix gliner_config.json: set vocab_size=128100 to match saved weights"). After that fix the model presumably loads, but there is nothing useful to load.

Files

Intermediate pushes (steps 500 … 17,000, every 500) are in the git history. Given the zero loss, none of them is expected to be better. There is no trainer state or metric in this repo to pick a "best" one. attempt_vanilla keeps gliner_logs/h-gliner-token-full-v1/checkpoint-17000/ (config + trainer state only).

Training data

Same as the large run: in-house synthetic PII, data/train.json (~1,026,320 examples) and data/eval.json. Not published.

Training recipe

From gliner_config.json / training_args.bin (inspected without unpickling) and configs/vanilla_h.yaml:

• Encoder microsoft/deberta-v3-small (6 layers, hidden 768), no GLiNER initialisation (prev_path: "none").
• span_mode: token_level, max_len 768, max_width 100, max_types 25, hidden_size 768, dropout 0.1, 1 RNN layer, class_token_index 128002.
• AdamW (torch); LR encoder 1e-5, LR others 5e-5; weight decay 0.01; cosine schedule, 10% warmup; max_steps 60000; batch 96 (eval 88); grad clip 1.0; bf16; gradient checkpointing; seed 42.
• Loss: focal_loss_alpha -1 (focal weighting off), gamma 0, sum reduction; negatives 2; masking: global.

Evaluation

• An April-2026 version of this card included a cross-model "real nervaluate" table on the private PT-BR holdout arthrod/pii-gliner-evals. In it, h-gliner-token-full-v1 scored 0.0000 on every column (Strict/Partial/Exact/EntType F1, P, R).
• Training-time loss was 0.0 throughout (see attempt_vanilla), so no learning signal was recorded.

Limitations

• Non-functional. Kept only as a record of a failed run.
• Lesson recorded in attempt_vanilla: assert loss > 0 and grad_norm > 0 in the first few dozen steps.

License

Apache-2.0 as declared. The encoder microsoft/deberta-v3-small is MIT.