---
license: apache-2.0
language:
- ru
base_model:
- ai-forever/ruBert-large
tags:
- difficulty
- cefr
- regression
---
# Model Card for Model ID

Regression model which predicts difficulty score for an input text. Predicted scores can be mapped to CEFR levels. 


## Model Details

Frozen BERT-large layers with a regressor on top. Trained on a mix of manually annotated datasets (more details on data will follow).



## How to Get Started with the Model

Use the code below to get started with the model.

```
class CustomModel(BertPreTrainedModel):
    def __init__(self, config, load_path=None, use_auth_token: str = None,):
        super().__init__(config)
        self.bert = BertModel(config)
        self.pre_classifier = nn.Linear(config.hidden_size, 128)
        self.dropout = nn.Dropout(0.2)
        self.classifier = nn.Linear(128, 1)

        # Apply Xavier initialization
        nn.init.xavier_uniform_(self.pre_classifier.weight)
        nn.init.xavier_uniform_(self.classifier.weight)
        if self.pre_classifier.bias is not None:
            nn.init.constant_(self.pre_classifier.bias, 0)
        if self.classifier.bias is not None:
            nn.init.constant_(self.classifier.bias, 0)

    
    def forward(
            self,
            input_ids,
            labels=None,
            attention_mask=None,
            token_type_ids=None,
            position_ids=None,
    ):
        outputs = self.bert(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
        )


        pooled_output = outputs[0][:, 0]
        pooled_output = self.pre_classifier(pooled_output)
        pooled_output = nn.ReLU()(pooled_output)  
        pooled_output = self.dropout(pooled_output)  
        logits = self.classifier(pooled_output) 

        if labels is not None:
            loss_fn = nn.MSELoss()
            loss = loss_fn(logits.view(-1), labels.view(-1))
            return loss, logits
        else:
            return None, logits


tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
config.num_labels = 1

model = CustomModel(config)
model.load_state_dict(torch.load(f'{model_path}/pytorch_model.bin'))

inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)
inputs = {key: value.to(device) for key, value in inputs.items()}
        
with torch.no_grad():
  _, logits = model(input_ids=inputs["input_ids"], attention_mask=inputs["attention_mask"], token_type_ids=inputs["token_type_ids"])

```

To map to CEFR, use: 
```
reg2cl2 = {'1.0': 'A1', '1.5': 'A12', '2.0': 'A2', '2.5': 'A2', '3.0': 'B1', '3.5': 'B12', '4.0': 'B2', '4.5': 'B2', '5.0': 'C1', '5.5': 'C12', '6.0': 'C2', '0.0': 'A1'}
print("Predicted output (logits):", logits.item(), reg2cl2[str(float(round(logits.item())))])
```



## Training Details


#### Training Hyperparameters

+ learning_rate: 3e-4
+ num_train_epochs: 15.0
+ batch_size: 32
+ weight_decay: 0.1
+ adam_beta1: 0.9
+ adam_beta2: 0.99
+ adam_epsilon: 1e-8
+ max_grad_norm:  1.0
+ fp16: True



## Evaluation on test set 


![Evaluation results](ru_regression.png)

## Citation

Please refer to this repo when using the model.