arthrod/gliner_review_comparison

GLiNER PII review comparison: 18 models × 8 eval sets

Per-sample predictions of 18 GLiNER checkpoints (the author's PT-BR PII fine-tunes plus public knowledgator, urchade and Ihor models) on 8 NER/PII eval sets, with one config per eval set and one split per model, so you can open a single model's outputs on a single eval set in the viewer. 904,952 prediction rows plus a 148-row summaries table (precision / recall / F1 and a per-label breakdown per model × eval set). Produced on 26–27 Feb 2026 in one evaluation sweep on an AMD MI300X.

Related repos

repo	type	role
arthrod/gliner_eval_folder (private)	dataset	the same Feb 2026 sweep, organised per eval set (all models in one split), plus the gold eval_sets, eval harness and logs
arthrod/attempt_vanilla	model	Feb 2026 GLiNER PII fine-tune experiment log (DeBERTa-v3-large from knowledgator/gliner-pii-large-v1.0); source of the gliner-pii-large-full-v1/checkpoint-* rows
arthrod/v1-deberta_large-token_level-1-7	model	evaluated here as arthrod/gliner-pii-large-full-v1 (its former name; the data keeps the old id); best generaliser among the fine-tunes
arthrod/gliner-opf-ptbr-pii-bench-v1	dataset	later (Apr 2026) benchmark: OpenAI Privacy Filter fine-tune vs GLiNER mmBERT/ettin
arthrod/gliner-opf-ptbr-pii-v1, arthrod/gliner-mmbert-small-ptbr-pii-full-3x-v1, arthrod/gliner-ettin-32m-ptbr-pii-full-3x-v1, arthrod/gliner-ettin-68m-ptbr-pii-full-3x-v1, gliner-ettin-68m-ptbr-pii-full-3x-v1-checkpoints/gliner-ettin-68m-ptbr-pii-top50k-v1	model	later PT-BR PII models of the same family (not evaluated here)
arthrod/gliner-opf-ptbr-pii-demo	Space	demo of the family's PT-BR PII detector

Configs and splits

config	eval set	samples	models (splits)
summaries (default)	all, one row per model × eval set (split summaries)	148 rows	—
ptbr_pii_val_19k	author's synthetic PT-BR PII validation set (val.json, ≤384 tokens)	19,164	18
canonical_ptbr_pii_v1	arthrod/gliner-canonical-ptbr-pii-v1 (synthetic PT-BR; private repo)	5,000	18
nemotron_pii	nvidia/Nemotron-PII	5,000	18
synthetic_pii_ner_mistral_v1	urchade/synthetic-pii-ner-mistral-v1	5,000	18
gliner_multilingual_synthetic	knowledgator/gliner-multilingual-synthetic	5,000	14
gliner_multi_task_synthetic_data	knowledgator/GLINER-multi-task-synthetic-data	5,000	18
pile_mistral_v0_1	urchade/pile-mistral-v0.1	5,000	18
gliner_post_train	Ihor/gliner-post-train	5,000	8

Split names are the model ids with / → __ and - → _: Ihor__gliner_pii_small, arthrod__gliner_lgpd_br_tokenlevel_lora_small_v1, arthrod__gliner_lgpd_br_tokenlevel_lora_small_v1_15k_fixed, arthrod__gliner_lgpd_br_tokenlevel_lora_small_v1_20k_fixed, arthrod__gliner_multitask_v1.0, arthrod__gliner_pii_large_full_v1, arthrod__gliner_pii_mmbert_base_token_v1.0, gliner_pii_large_full_v1__checkpoint_1000, gliner_pii_large_full_v1__checkpoint_8000, knowledgator__gliner_linker_base_v1.0, knowledgator__gliner_linker_large_v1.0, knowledgator__gliner_linker_rerank_v1.0, knowledgator__gliner_multitask_large_v0.5, knowledgator__gliner_multitask_v1.0, knowledgator__gliner_pii_large_v1.0, knowledgator__gliner_pii_small_v1.0, urchade__gliner_instruct_beta, urchade__gliner_token_medium_v1. Not every model was run on every set (see the results table; a dash = no split).

The gliner-pii-large-full-v1/checkpoint-* models are local checkpoints of the attempt_vanilla run, evaluated before upload (their model field is the local path).

Schema

Prediction configs (all except summaries)

column	type	description
sample_idx	int32	index of the sample within the eval set (same across splits of a config)
model	string	model id (or local checkpoint path) that produced the predictions
tokens	list	the tokenised input text
ground_truth	list<struct{start: int32, end: int32, label: string}>	gold spans; start/end are inclusive token indices
predictions	list<struct{start: int32, end: int32, label: string, score: float64}>	predicted spans (inclusive token indices) with model confidence

summaries

column	type	description
eval_dataset	string	eval set id
config	string	prediction config holding this row's samples (null for the invalid pubmed rows)
model	string	model id as logged (local path for checkpoints)
model_name_clean	string	short model name
split	string	split name inside config (null for pubmed rows)
threshold	float64	0.5 for every run
num_samples	int32	samples scored
precision, recall, f1	float64	micro-averaged exact-span scores
load_time, eval_time	float64	seconds
known_issue	string	null, or a note for rows that should not be trusted (see Caveats)
source_folder, source_file	string	raw file the row came from (inference_results*/<model>.jsonl, now removed)
per_entity	list<struct{label, precision, recall, f1, tp: int32, pred_count: int32, true_count: int32}>	per-label breakdown

Examples

ptbr_pii_val_19k, split arthrod__gliner_pii_large_full_v1:

{"sample_idx": 0, "model": "arthrod/gliner-pii-large-full-v1",
 "tokens": ["Oi", ",", "Talison", ".", "Tô", "de", "saco", ".", "Meu", "número", "é", "(", "51)", "5722-6509", ",", "liga", ".", "Sou", "da", "Silva", "Soares", ",", "solteiro", ",", "não", "…"],
 "ground_truth": [{"start": 63, "end": 63, "label": "email address"}, {"start": 2, "end": 2, "label": "first name"}, {"start": 18, "end": 20, "label": "last name"}, "…"],
 "predictions": [{"start": 2, "end": 2, "label": "location city", "score": 0.5666}, {"start": 9, "end": 9, "label": "location building number", "score": 0.9740}, {"start": 22, "end": 41, "label": "subject described sexual data", "score": 0.9756}, "…"]}

nemotron_pii, split knowledgator__gliner_pii_small_v1.0:

{"sample_idx": 0, "model": "knowledgator/gliner-pii-small-v1.0",
 "tokens": ["You", "are", "cordially", "invited", "to", "our", "upcoming", "agricultural", "event", "titled", "\"", "Harvesting", "the", "Future", ".\"", "The", "event", "will", "take", "place", "on", "Fri", ",", "15", "Aug", "…"],
 "ground_truth": [{"start": 21, "end": 25, "label": "date"}, {"start": 29, "end": 29, "label": "time"}, {"start": 32, "end": 32, "label": "city"}, "…"],
 "predictions": [{"start": 21, "end": 25, "label": "date", "score": 0.7099}, {"start": 29, "end": 29, "label": "time", "score": 0.7372}, {"start": 32, "end": 32, "label": "city", "score": 0.7399}, "…"]}

summaries:

{"eval_dataset": "arthrod ptbr_pii_val_19k (val.json)", "config": "ptbr_pii_val_19k", "model": "Ihor/gliner-pii-small",
 "model_name_clean": "Ihor/gliner-pii-small", "split": "Ihor__gliner_pii_small", "threshold": 0.5, "num_samples": 19164,
 "precision": 0.0814, "recall": 0.1866, "f1": 0.1133, "load_time": 1.83, "eval_time": 144.27, "known_issue": null,
 "source_folder": "inference_results", "source_file": "Ihor__gliner-pii-small.jsonl",
 "per_entity": [{"label": "AS", "precision": 0.0, "recall": 0.0, "f1": 0.0, "tp": 0, "pred_count": 0, "true_count": 3}, "…"]}

How it was built

The sweep ran on 26–27 Feb 2026 on one AMD MI300X with the eval harness published in arthrod/gliner_eval_folder (private) (run_eval.py, prepare_datasets.py). Up to 5,000 samples per eval set (all 19,164 for the PT-BR val set); each eval set is scored with its own gold label set (capped at 25 labels per batch), threshold 0.5. Each run wrote one JSONL per model and eval set: a summary line followed by one sample line per input. Those files (inference_results*/<model>.jsonl, 1.9 GB) were originally uploaded here together with a flattened summaries parquet.

Parquet conversion (Oct 2026). Every JSONL was converted to data/<eval set>/<model>.parquet (sample lines) and its summary line to a row of data/summaries/summaries.parquet (per_entity_json string → typed per_entity list; num_samples → int). The 148 summary rows match the previous summaries parquet exactly (precision/recall/F1 compared row by row). The raw JSONL files were then removed. The 18 pubmed-ner-mistral-v0.1 files were not converted to predictions (see Caveats); their summary rows are kept with known_issue set.

Usage

from datasets import load_dataset

summ = load_dataset("arthrod/gliner_review_comparison", "summaries", split="summaries").to_pandas()
preds = load_dataset("arthrod/gliner_review_comparison", "nemotron_pii", split="arthrod__gliner_pii_large_full_v1")

# exact-span hits for one sample
row = preds[0]
gold = {(g["start"], g["end"], g["label"]) for g in row["ground_truth"]}
hits = [p for p in row["predictions"] if (p["start"], p["end"], p["label"]) in gold]

Results (exact-span F1, threshold 0.5) — summaries config

Sorted by the mean over the first four (PII) columns. A dash means the model wasn't run on that set.

model	ptbr_pii_val_19k	canonical_ptbr_pii_v1	nemotron_pii	synthetic_pii_ner_mistral_v1	gliner_multilingual_synthetic	gliner_multi_task_synthetic_data	pile_mistral_v0_1	gliner_post_train
arthrod/gliner-pii-large-full-v1	0.383	0.730	0.627	0.246	0.113	0.081	0.129	0.131
gliner-pii-large-full-v1/checkpoint-8000	0.385	0.730	0.615	0.246	0.113	0.081	0.129	0.131
gliner-pii-large-full-v1/checkpoint-1000	0.367	0.691	0.601	0.263	0.121	0.096	0.133	—
arthrod/gliner-lgpd-br-tokenlevel-lora-small-v1-15k-fixed	0.708	0.666	0.443	0.084	0.420	0.091	0.163	0.136
arthrod/gliner-lgpd-br-tokenlevel-lora-small-v1-20k-fixed	0.713	0.667	0.440	0.071	0.423	0.089	0.158	0.130
arthrod/gliner-lgpd-br-tokenlevel-lora-small-v1	0.713	0.668	0.436	0.073	0.428	0.090	0.159	0.132
arthrod/gliner-pii-mmbert-base-token-v1.0	0.713	0.668	0.436	0.073	0.428	0.090	0.159	—
arthrod/gliner-multitask-v1.0	0.322	0.615	0.513	0.223	—	0.167	0.131	0.140
knowledgator/gliner-pii-large-v1.0	0.183	0.459	0.581	0.283	0.135	0.190	0.157	—
Ihor/gliner-pii-small	0.113	0.273	0.562	0.538	0.101	0.011	0.207	—
knowledgator/gliner-multitask-v1.0	0.137	0.384	0.578	0.269	—	0.195	0.161	0.160
knowledgator/gliner-pii-small-v1.0	0.070	0.187	0.621	0.451	—	0.087	0.161	—
knowledgator/gliner-multitask-large-v0.5	0.155	0.368	0.530	0.269	0.147	0.200	0.136	—
knowledgator/gliner-linker-rerank-v1.0	0.091	0.200	0.570	0.312	—	0.078	0.134	—
urchade/gliner_token_medium-v1	0.128	0.352	0.422	0.212	0.153	0.003	0.150	0.123
knowledgator/gliner-linker-large-v1.0	0.064	0.254	0.209	0.091	0.041	0.164	0.063	—
urchade/gliner_instruct-beta	0.075	0.203	0.213	0.120	0.079	0.002	0.058	—
knowledgator/gliner-linker-base-v1.0	0.045	0.120	0.180	0.080	0.043	0.129	0.055	—

Takeaways:

• The PT-BR LoRA/token-level fine-tunes win the in-domain PT-BR val set (~0.71 vs ≤0.18 for off-the-shelf models) but drop to 0.07–0.08 on urchade's English synthetic PII: they overfit to their training distribution.
• gliner-pii-large-full-v1 generalises best among the fine-tunes: 0.73 on PT-BR canonical and 0.63 on Nemotron-PII (the best Nemotron score in the sweep), but only 0.38 on the 19k PT-BR val set.
• Off-the-shelf small PII models (Ihor/gliner-pii-small, knowledgator/gliner-pii-small-v1.0) are strongest on the English synthetic PII sets and weak on PT-BR.

Limitations: caveats and known issues

• The pubmed results were invalid and are not published as predictions. The 18 files under inference_results_urchade__pubmed-ner-mistral-v0.1/ were byte-identical to the pile-mistral-v0.1 files (pile sentences, pile scores): the harness's preprocessing cache served pile batches for pubmed. The 18 pubmed rows remain in summaries with config = null and a known_issue note; do not use their numbers.
• arthrod/gliner-pii-mmbert-base-token-v1.0 and arthrod/gliner-lgpd-br-tokenlevel-lora-small-v1 produced byte-identical predictions on all 7 eval sets they share. Either the two Hub repos hold the same weights or the eval mixed them up; treat the pair as one result. Their summaries rows carry a known_issue note.
• arthrod/gliner-pii-large-full-v1 and gliner-pii-large-full-v1/checkpoint-8000 give identical sample-level predictions on 6 of the 8 sets (the final model is that checkpoint); their summary F1 differs slightly on ptbr_pii_val_19k and nemotron_pii.
• Label sets differ per eval set (each uses its own gold labels, capped at 25 per batch). Scores are comparable across models within a column, not across columns. ptbr_pii_val_19k mixes PT-BR PII labels with multilingual general NER labels.
• The PT-BR sets contain only synthetic PII. Nemotron-PII and the PT-BR sets deliberately include fake API keys, JWTs and test tokens (e.g. Stripe's public docs test key) because "api key" is a PII label; none are real credentials.
• Several arthrod/* models in this sweep (gliner-lgpd-br-tokenlevel-lora-small-v1*, gliner-multitask-v1.0, gliner-pii-mmbert-base-token-v1.0) and the gliner-canonical-ptbr-pii-v1 eval set are private repos.

License & attribution

No single license: predictions inherit the terms of the eval set they were computed on. nvidia/Nemotron-PII is CC-BY-4.0; urchade/synthetic-pii-ner-mistral-v1, urchade/pile-mistral-v0.1 and knowledgator/GLINER-multi-task-synthetic-data are Apache-2.0; knowledgator/gliner-multilingual-synthetic and Ihor/gliner-post-train declare no license on the Hub — check with their authors before reuse. The author's PT-BR sets and the predictions/summaries themselves follow the Apache-2.0 terms used in arthrod/gliner_eval_folder (private). Models: GLiNER (Zaratiana et al., 2023) checkpoints from knowledgator, urchade and Ihor, plus the author's fine-tunes.