Базовая модель: ai-forever/ruRoBERTa-large
Задача: Multi-label классификация
Датасет: S7_enriched.jsonl
Теги (65): Fast Track, Gate 7, HR, Nordwind, S7 Cargo, S7 Engineering, S7 Technics, S7 Training, SMS информирование, Seats...
Best Micro F1: 0.8878 (epoch 15)
Пулинг: (CLS + mean_pool) / 2
Обучена: 2026-08-10

Инференс

python

import json, torch
from transformers import AutoTokenizer

with open("config.json") as f:
    cfg = json.load(f)

all_tags = cfg["all_tags"]
tokenizer = AutoTokenizer.from_pretrained("DanielNRU/S7-tags-ruroberta-20260810")
# model = RuRoBERTaTagClassifier("ai-forever/ruRoBERTa-large", num_tags=len(all_tags))
# model.load_state_dict(torch.load("best_pytorch_model.bin", map_location="cpu"))
model.eval()

text = "Потеряли багаж, служба поддержки не отвечает"
enc  = tokenizer(
    text,
    max_length=512,
    truncation=True,
    padding="max_length",
    return_tensors="pt",
)
with torch.no_grad():
    logits = model(enc["input_ids"], enc["attention_mask"])
    probs  = torch.sigmoid(logits).squeeze(0).tolist()

result = {
    all_tags[i]: round(probs[i], 3)
    for i in range(len(all_tags))
    if probs[i] >= 0.3
}
print(result)