matura-lol/GLiNER2.5-multi-Decide-finetune

GLiNER 2.5 Multi-Decide Fine-Tune (Round 2 - SemIf Distillation)

Multi-task classification of Polish exam questions — a LoRA fine-tune of GLiNER2.5-multi-Decide on the matura.lol datasets (Jev & SemIf tags), by matura.lol.

Trained with authentic, verified teacher tags generated by SemIf-OpenJev (Qwen3.5-4B) alongside Jev (jev-1.13) on the matura.lol corpus (matematyka, fizyka, polski). Zero pseudo-labels; zero unverified tags.

The model classifies a Polish exam question along four dimensions in a single forward pass:

Task	Description	Labels
topic	Subject-specific topic (per CKE informatory)	Varies by subject
method	Solution method	fakty, analiza_zrodla, rachunek, interpretacja_danych, wypowiedz, gramatyka, algorytm, inne, doswiadczenie, dowod, rysunek_techniczny
answer_form	Expected answer form	krotka_odpowiedz, wybor, liczba, wypracowanie, prawda_falsz, wyrazenie, dowod, rysunek, inne, kod
difficulty	Perceived difficulty	łatwe, bardzo łatwe, średnie, trudne

Evaluation Benchmark (Held-out validation split)

Performance on the held-out validation split after SemIf distillation:

Task	Accuracy	Macro-F1	Notes
answer_form	95.56%	19.77%	Highest accuracy across structural forms
method	90.00%	24.14%	Strong classification across solution methods
difficulty	63.33%	35.90%	Subjective difficulty assessment
Overall	82.96%	—	Average task accuracy across evaluation split

• Zero pseudo-labels; zero unverified tags.
• Substantial improvement over the initial baseline (~55.2% overall accuracy).

Model details

• Base model: fastino/GLiNER2.5-multi-Decide (287M, multilingual)
• Fine-tuning: LoRA (r=16, alpha=32, dropout=0.05) on the encoder and classifier heads, ~2.7M trainable params (0.94% of the model)
• Teacher models / Supervision: SemIf-OpenJev (Qwen3.5-4B) and Jev (jev-1.13) providing verified teacher tags
• Training data: Questions from the matura.lol corpus (matematyka, fizyka, polski), paired with verified Jev and SemIf-OpenJev tags
• License: AGPL-3.0 (see below)
• Organizations: github.com/matura-lol · huggingface.co/matura-lol

License

This model is released under the GNU Affero General Public License v3.0 (AGPL-3.0), matching the matura.lol datasets.

The base model fastino/GLiNER2.5-multi-Decide is Apache-2.0, which is permissive and permits relicensing derived works under a different (including copyleft) license. The fine-tune itself — including this adapter and the training setup — is distributed under AGPL-3.0, consistent with the rest of the matura.lol open-source project.

Usage

from gliner2 import AutoExtractor
from peft import PeftModel

base = AutoExtractor.from_pretrained("fastino/GLiNER2.5-multi-Decide", map_location="cpu")
model = PeftModel.from_pretrained(base, "matura-lol/GLiNER2.5-multi-Decide-finetune")

METHODS = ["fakty", "analiza_zrodla", "rachunek", "interpretacja_danych",
           "wypowiedz", "gramatyka", "algorytm", "inne", "doswiadczenie",
           "dowod", "rysunek_techniczny"]
FORMS = ["krotka_odpowiedz", "wybor", "liczba", "wypracowanie", "prawda_falsz",
         "wyrazenie", "dowod", "rysunek", "inne", "kod"]
DIFFICULTY = ["łatwe", "bardzo łatwe", "średnie", "trudne"]

schema = {
    "method": {"labels": METHODS},
    "answer_form": {"labels": FORMS},
    "difficulty": {"labels": DIFFICULTY},
}

pred = model.classify_text("Oblicz pole koła o promieniu 5.", schema)
print(pred)
# {"method": "rachunek", "answer_form": "liczba", "difficulty": "łatwe"}

Attribution

If you reuse this model or the underlying data — including for AI training or retrieval-augmented generation — credit matura.lol with a link. The dataset is AGPL-3.0; individual pieces of text may carry their own licences (CKE/OKE materials, third-party solutions).

Links

• Site: matura.lol
• GitHub org: github.com/matura-lol
• Hugging Face org: huggingface.co/matura-lol
• Dataset: matura.lol datasets
• Categorisation pipeline: Jev-categorise
• Base model: fastino/GLiNER2.5-multi-Decide
• Base paper: GLiNER: Generalist Model for Named Entity Recognition using Bidirectional Transformer · GLiNER2.5