README.md

# Multi-Task NER + Intent + Language Model

This model performs three tasks simultaneously:
1. **Named Entity Recognition (NER)**: Extracts entities from B2B transaction descriptions
2. **Intent Classification**: Classifies transaction intent/purpose
3. **Language Detection**: Detects language (English, Russian, Uzbek Latin/Cyrillic, Mixed)

## Model Details
- Base model: `google-bert/bert-base-multilingual-uncased`
- Architecture: Enhanced multi-task model with BiLSTM for NER, attention pooling for classification
- Training: Optimized for realistic B2B transaction descriptions

## Supported Languages
- English (en)
- Russian (ru)
- Uzbek Latin (uz_latn)
- Uzbek Cyrillic (uz_cyrl)
- Mixed language text

## Usage

```python
import torch
import torch.nn as nn
import numpy as np
import json
from transformers import AutoTokenizer, AutoModel
from huggingface_hub import hf_hub_download

# Download model files
model_id = "primel/aibanov"
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Download label mappings
mappings_file = hf_hub_download(repo_id=model_id, filename="label_mappings.json")
with open(mappings_file, "r") as f:
    label_mappings = json.load(f)

id2tag = {int(k): v for k, v in label_mappings["id2tag"].items()}
id2intent = {int(k): v for k, v in label_mappings["id2intent"].items()}
id2lang = {int(k): v for k, v in label_mappings["id2lang"].items()}

# Define model architecture (same as training)
class EnhancedMultiTaskModel(nn.Module):
    # ... (copy the model class from training script)
    pass

# Load model
base_bert = "google-bert/bert-base-multilingual-uncased"
model = EnhancedMultiTaskModel(
    model_name=base_bert,
    num_ner_labels=len(label_mappings["tag2id"]),
    num_intent_labels=len(label_mappings["intent2id"]),
    num_lang_labels=len(label_mappings["lang2id"]),
    dropout=0.15
)

# Load trained weights
weights_file = hf_hub_download(repo_id=model_id, filename="pytorch_model.bin")
state_dict = torch.load(weights_file, map_location='cpu')
model.load_state_dict(state_dict)
model.eval()

# Inference
text = "Оплата 100% за товары согласно договору №123 от 15.01.2025г ИНН 987654321"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=192)

with torch.no_grad():
    outputs = model(**inputs)

# Process outputs
ner_logits = outputs['ner_logits'][0].numpy()
intent_logits = outputs['intent_logits'][0].numpy()
lang_logits = outputs['lang_logits'][0].numpy()

# Get predictions
intent_id = np.argmax(intent_logits)
intent = id2intent[intent_id]
print(f"Intent: {intent}")

lang_id = np.argmax(lang_logits)
language = id2lang[lang_id]
print(f"Language: {language}")
```

## License
Apache 2.0

## Citation
```bibtex
@misc{aibanov2025,
  author = {primel},
  title = {Multi-Task NER Intent Language Model},
  year = {2025},
  url = {https://huggingface.co/primel/aibanov}
}
```