kiselyovd/grnti-text-classifier

Production-grade Russian scientific-text classifier: 28 top-level GRNTI codes. Main model: XLM-RoBERTa-base fine-tuned on ai-forever/ru-scibench-grnti-classification.

Metrics (test split, n = 2772, 28 classes)

Model	Top-1	Top-5	Macro F1	Weighted F1
FacebookAI/xlm-roberta-base	72.4%	96.8%	72.3%	72.3%
DeepPavlov/rubert-base-cased	72.9%	95.9%	72.8%	72.8%

Usage

from transformers import pipeline

clf = pipeline("text-classification", model="kiselyovd/grnti-text-classifier", top_k=5)
clf("Исследование квантовой электродинамики в кристаллах.")

Intended use

This model is trained for Russian-language top-level GRNTI section classification (State Rubricator of Scientific and Technical Information). It is not evaluated outside Russian scientific text and should not be used for generic multilingual classification.

Do not rely on this model for high-stakes decisions. Outputs are probabilistic and subject to training-data biases.

Training

• Dataset: ai-forever/ru-scibench-grnti-classification (MIT, 28 476 train + 2 772 test).
• Base model: FacebookAI/xlm-roberta-base.
• Baseline: DeepPavlov/rubert-base-cased.
• Precision: bf16-mixed on CUDA.
• Optimizer: AdamW + linear warmup/decay.
• Optuna 10-trial sweep for lr/weight_decay/warmup_ratio, then 5-epoch final training with best params.

Source: https://github.com/kiselyovd/grnti-text-classifier

License

MIT.