Update how_to_use.md

how_to_use.md

@@ -126,7 +126,9 @@ For model inference you will need to execute the same steps on new input text.
 
 ---
 
-## Multitask Model definition
+## Multitask Model
+
+### 1. Multitask Model class
 To use the model you will need to define a `multitask_model.py` with the custom model class built upon our base model.
 
 ```python
@@ -205,4 +207,82 @@ class MultitaskModel(PreTrainedModel):
             "intent_logits": intent_logits,
             "ner_logits": ner_logits,
         }
-```
+```
+
+---
+
+### 2. Load Tokenizer
+
+```
+model_name = "distilbert-base-uncased"
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+```
+
+---
+
+### 3. Custom Metrics Function
+This function is essential for a multitask model. It will be called by the Trainer at the end of each epoch to calculate both intent accuracy and NER F1-score.
+
+```python
+def compute_metrics(eval_pred):
+    # Unpack predictions and labels
+    predictions, label_values = eval_pred
+    intent_preds, ner_preds = predictions
+    intent_labels, ner_labels = label_values
+
+    # --- Intent Metrics ---
+    intent_preds = np.argmax(intent_preds, axis=1)
+    intent_accuracy = accuracy_score(intent_labels, intent_preds)
+    intent_f1 = f1_score(intent_labels, intent_preds, average='weighted')
+
+    # --- NER Metrics ---
+    ner_preds = np.argmax(ner_preds, axis=2)
+
+    # Remove padding tokens (where label is -100) and convert IDs to labels
+    true_ner_labels = []
+    true_ner_predictions = []
+    id2ner = processed_datasets['train'].features['labels'].feature.names
+
+    for i in range(len(ner_labels)):
+        true_labels_row = []
+        true_predictions_row = []
+        for j in range(len(ner_labels[i])):
+            if ner_labels[i][j] != -100:
+                true_labels_row.append(id2ner[ner_labels[i][j]])
+                true_predictions_row.append(id2ner[ner_preds[i][j]])
+        true_ner_labels.append(true_labels_row)
+        true_ner_predictions.append(true_predictions_row)
+
+    ner_f1 = ner_f1_score(true_ner_labels, true_ner_predictions, mode='strict', scheme=IOB2)
+
+    return {
+        "intent_accuracy": intent_accuracy,
+        "intent_f1": intent_f1,
+        "ner_f1": ner_f1
+    }
+```
+
+---
+
+### 4. Instantiate the model
+We now create an instance of our `MultitaskModel`, passing it a configuration object that includes the number of labels for each head.
+
+```
+# Get label mappings from the dataset features
+id2intent = processed_datasets['train'].features['intent_label'].names
+intent2id = {name: i for i, name in enumerate(id2intent)}
+id2ner = processed_datasets['train'].features['labels'].feature.names
+ner2id = {name: i for i, name in enumerate(id2ner)}
+
+# Load the model config and add our custom parameters
+config = AutoConfig.from_pretrained(
+    model_name,
+    id2label_intent=id2intent,
+    label2id_intent=intent2id,
+    id2label_ner=id2ner,
+    label2id_ner=ner2id
+)
+
+# Instantiate our custom model with the new config
+model = MultitaskModel(config, num_intent_labels=len(id2intent), num_ner_labels=len(id2ner))
+```