Intento-v1-xlmr

Intento-v1

Intento-v1 is a fine-tuned XLM-RoBERTa base model optimized for high-accuracy multilingual intent classification. It was trained on the SID dataset. It supports 101 distinct intents across 5 languages (English, Portuguese, Spanish, French, and German).

Model Details

• Author: luigicfilho
• Version: 1.0.0
• Base Model: xlm-roberta-base
• Training Task: Multilingual Intent Classification
• Accuracy: 95.79% on validation set.

Files

• checkpoint.pt: Full 3.11 GB PyTorch checkpoint including model weights, optimizer state (AdamW), and training metadata.
• labels.json: Full mapping of the 101 intent classes.
• model_config.json: Core model configuration tokens.

Intent Classes

This model classifies text into 101 distinct intents. The labels are sorted alphabetically and can be found in the attached labels.json.

How to Use

This model is intended to be loaded directly from the checkpoint.pt file. This preserves the full training state, including the optimizer and metadata.

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

# 1. Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-base")

# 2. Load the full checkpoint
checkpoint_path = "checkpoint.pt"
ckpt = torch.load(checkpoint_path, map_location="cpu")

# 3. Initialize the model structure
# We use xlm-roberta-base with 101 labels matching the training config
model = AutoModelForSequenceClassification.from_pretrained(
    "xlm-roberta-base", 
    num_labels=101
)

# 4. Load the weights from the checkpoint dictionary
model.load_state_dict(ckpt['model_state_dict'])
model.eval()

# 5. Inference Example
text = "I need help with my account"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=64)

with torch.no_grad():
    logits = model(**inputs).logits
    predicted_class_id = logits.argmax().item()

print(f"Predicted class ID: {predicted_class_id}")
print(f"Model Metadata: {ckpt.get('metadata')}")

License

This model is licensed under the MIT License.