gliner2.5-mursit-kvkk-tr-v1

A GLiNER2 extractor on NewMind's Turkish ModernBERT encoder (Mursit-Base) instead of mDeBERTa: 165M parameters, 3–5× faster at inference. Trained in two stages — 100,000 real Turkish legal and KVKK sentences (30 labels, entities only), then nm-kvkk-pii-6K with Turkish names where every training record carries the complete 113-type relation schema, which is what makes this encoder work with long prompts.

Part of the nm-kvkk release (https://github.com/newmindai/nm-kvkk): training recipes, the patched GLiNER2 library, the synthetic-data generator and the evaluators behind every number below.

Model details

Architecture	GLiNER2 boundary extractor: schema-prompted entities + relations in one pass
Encoder	Mursit-Base — Turkish ModernBERT, 165M, positional range extended from 1,024 to 4,096 tokens; GLiNER2.5 span and relation heads warm-started from fastino/gliner2.5-multi-v1
Label language	Turkish label names
Taxonomy	KVKK personal-data taxonomy v2.2 — 118 entity types, 105 relation types (vocabulary files in the repository)
Licence	Apache-2.0

Training recipe

Encoder	newmindai/Mursit-Base with max_position_embeddings 1,024 → 4,096 and NTK-scaled global RoPE (θ 10,000 → 41,829); native byte-level tokenization, mean pooling over sub-words; heads copied from gliner2.5-multi-v1 (136/136 tensors)
Stage 1	MIX100k: 100,000 real Turkish legal + KVKK sentences, 30 labels, entities only; 13,000 steps, batch 4 on 1×H100, lr 1e-5 / 5e-4, constant with 20 warm-up steps
Stage 2	nm-kvkk-pii-6K with Turkish names and the full 113-type relation schema in every record; 13,000 steps, same optimiser; documents capped at 1,024 words
Precision	bf16; weight decay 0.01; seed 42
Label language	Turkish entity and relation names
Inference prompt	entity labels + relation types as bare names (no descriptions) — this model loses recall with long prompts; whole document in one pass

Use bare relation names, not descriptions. Same real-document entity score as the c2000 model on the broad set at a third of the latency; weaker relations and lower recall than the mDeBERTa models.

How to use

This model requires the GLiNER2 library shipped in the nm-kvkk repository, not the PyPI package. Its encoder is a byte-level ModernBERT and its configuration carries fields the upstream library does not know (tokenization: native, add_bos, token_pooling: mean). With pip install gliner2 (2.0.0) the checkpoint loads without any error and returns wrong, incomplete output — the fields are silently ignored. Install the library from the repository instead:

git clone https://github.com/newmindai/nm-kvkk.git && cd nm-kvkk
pip install -r requirements.txt
export PYTHONPATH=$PWD/GLiNER2   # or: pip install -e GLiNER2

Out of the box — extract_kvkk.py ships in this repository: text in, JSON keyed by taxonomy ids out, the label names handled for you.

pip install gliner2 huggingface_hub
python extract_kvkk.py --model newmindai/gliner2.5-mursit-kvkk-tr-v1 document.txt                   # 19 KVKK fields + person and company names
python extract_kvkk.py --model newmindai/gliner2.5-mursit-kvkk-tr-v1 --relations all document.txt   # ... plus every relation type
python extract_kvkk.py --model newmindai/gliner2.5-mursit-kvkk-tr-v1 --labels all --relations all document.txt   # the whole taxonomy
python extract_kvkk.py --model newmindai/gliner2.5-mursit-kvkk-tr-v1 --text "Ahmet Yılmaz, TC kimlik numarası 12345678901, ..."

(get the script with huggingface_hub.hf_hub_download("newmindai/gliner2.5-mursit-kvkk-tr-v1", "extract_kvkk.py") or from the repository.)

In your own code:

import json
from gliner2 import AutoExtractor
from huggingface_hub import hf_hub_download

repo = "newmindai/gliner2.5-mursit-kvkk-tr-v1"
model = AutoExtractor.from_pretrained(repo)

# kvkk_schema.json ships with the model: every entity label and relation type, with the exact query
# names this model was trained on, so you never need to know whether it expects Turkish or English names.
S = json.load(open(hf_hub_download(repo, "kvkk_schema.json"), encoding="utf-8"))
ent_name = {e["id"]: e["name"] for e in S["entities"]}          # taxonomy id -> query name
rel_name = {r["id"]: r["name"] for r in S["relations"]}
rel_desc = {r["name"]: r["description"] for r in S["relations"]}

# ask for what you need — a shorter prompt is faster and, for these models, more accurate
wanted = S["subsets"]["stack21"]                                   # or S["subsets"]["kvkk19"], or any list of ids
relations = ["national_id_of", "phone_of", "email_of"]             # ids from S["relations"]

schema = model.create_schema().entities([ent_name[i] for i in wanted])
if S["relation_prompt"] == "desc":
    schema = schema.relations({rel_name[r]: rel_desc[rel_name[r]] for r in relations})
else:                                                              # this model works best with bare relation names
    schema = schema.relations([rel_name[r] for r in relations])

out = model.extract(text, schema, threshold=0.5, include_confidence=True, include_spans=True)
out["entities"]            # {query name: [{text, start, end, confidence}]}
out["relation_extraction"] # {query name: [{head, tail}]}

# map the query names in the output back to taxonomy ids
name_to_id = {v: k for k, v in ent_name.items()}
entities = {name_to_id[n]: spans for n, spans in out["entities"].items()}

kvkk_schema.json fields: entities[] = {id, name, group, out_of_scope, trained}, relations[] = {id, name, description}, subsets (the 19 KVKK fields and the 21-label set used in the benchmarks), label_language and relation_prompt (the prompt shape this model works best with). The same ids are used by every model of the release, so switching model only changes the names the file maps them to. The full listing is in the section below. For this model, pass relation types as bare names without descriptions.

Benchmark results

All models below were run through the same pipeline and scored by the same code (see the repository). Scores are strict character-offset precision / recall / F1 (start, end and label must all match), micro over spans; macro F1 averages the per-label F1 over the labels present; lenient F1 accepts an overlapping span with the right label (boundary conventions differ between word-level taggers and span models). Every model is queried with the 21 labels all of them can express (19 KVKK fields + person name + company name); gold is restricted to those labels. Threshold 0.5 for the GLiNER2 models.

mixed_v2 — 118 real documents, 19 document families (1323 gold spans in the 21 shared labels, whole-document input)

model	P	R	F1	macro F1	lenient F1
gliner2.5-kvkk-tr-v2	76.6	76.9	76.8	61.3	81.3
gliner2.5-mursit-kvkk-tr-v1	84.8	59.2	69.7	54.6	73.6
YTU-ModernBERT-TR-PII	61.2	78.8	68.9	47.8	78.6
gliner2.5-kvkk-tr-v1	74.8	62.7	68.2	54.6	73.3
GLiNER2.5-multi (untuned)	65.7	55.3	60.0	44.4	66.5
redact (LiteRT)	36.0	22.0	27.3	14.1	39.2

vekaletname — 20 real notary documents (188 gold spans in the 21 shared labels, whole-document input)

model	P	R	F1	macro F1	lenient F1
gliner2.5-kvkk-tr-v2	91.3	89.4	90.3	93.9	93.0
gliner2.5-kvkk-tr-v1	90.4	80.3	85.1	86.4	89.6
gliner2.5-mursit-kvkk-tr-v1	91.2	66.0	76.5	80.6	79.6
YTU-ModernBERT-TR-PII	68.8	79.8	73.9	74.4	81.8
GLiNER2.5-multi (untuned)	73.5	53.2	61.7	60.5	76.5
redact (LiteRT)	28.9	11.7	16.7	27.9	28.8

*nm-kvkk-pii-6K test — 550 synthetic documents (in-distribution for models trained on nm-kvkk-pii-6K, marked ) (2289 gold spans in the 21 shared labels, whole-document input)

model	P	R	F1	macro F1	lenient F1
gliner2.5-mursit-kvkk-tr-v1*	92.9	88.4	90.6	92.1	96.3
gliner2.5-kvkk-tr-v2*	85.2	93.4	89.1	84.0	94.9
gliner2.5-kvkk-tr-v1	83.9	92.0	87.7	82.6	93.6
GLiNER2.5-multi (untuned)	67.9	65.0	66.4	51.3	76.7
YTU-ModernBERT-TR-PII	60.4	69.1	64.5	48.3	86.7
redact (LiteRT)	40.7	25.3	31.2	19.2	58.7

Full taxonomy and relation extraction

The same model queried with every label of the taxonomy it was trained with and every relation type, whole document in one pass — the way it is meant to be deployed. Relation F1 is strict on type and both end-points, coreference respected; the type-filtered column applies training-derived type constraints as a post-filter.

set	entity F1 (P / R)	lenient	relation F1 (P / R)	relation F1, type-filtered
vekaletname (20 real notary documents)	68.9 (73.1 / 65.1)	70.4	36.4 (28.1 / 51.9)	36.9

Limitations

• Trained on synthetic Turkish documents (plus, for the MIX100k models, real sentences); scores on real documents are 10–25 points below the synthetic test split. The real-document sets used here are small (20 and 118 documents).
• Company and institution names are the hardest label for every model in this family.
• Sentence-level input hurts: feed whole documents or paragraphs, not individual sentences.
• Pseudonymised real data was used for evaluation only; no real personal data was used in training.

Citation

NewMind AI, 2026 — nm-kvkk: Turkish KVKK personal-data extraction with GLiNER2.5, https://github.com/newmindai/nm-kvkk.