tferhan/finetuned_camb_intents

Model Card for Model ID

This model is fine-tuned from the camembert-base model and is designed to classify user intent questions for the website data.gov.ma in French. It can distinguish whether a user is making a general inquiry or requesting specific data. The training data was generated using GPT-4o-mini and includes information specific to data.gov.ma. The model was fine-tuned using LoRA with specific hyperparameters, achieving an accuracy of up to 0.98.

Model Details

Model Description

• Developed by: TFERHAN
• Language: French
• License: Apache 2.0
• Finetuned from model: camembert-base

Use Case

• Purpose: Classify user intent questions for the chatbot on the data.gov.ma website.
• Languages: French (optimized for), performs poorly on other languages.
• Data Source: Generated using GPT-4o-mini with data from data.gov.ma.

Uses

Direct Use

The model can be directly used to classify user intents in chatbot scenarios for the website data.gov.ma, distinguishing between general inquiries and data requests.

Downstream Use

The model is particularly suited for applications involving the French language and can be integrated into larger chatbot systems or fine-tuned further for similar tasks in different contexts.

Out-of-Scope Use

• Misuse for different languages without fine-tuning.
• Applications that do not involve French language queries.
• Sensitive or highly critical applications without extensive validation.

Bias, Risks, and Limitations

Technical Limitations

• Performance may degrade significantly on languages other than French.
• Limited to intents related to general queries and data requests.

Recommendations

• The model should be retrained or fine-tuned with appropriate data before deployment in non-French contexts.
• Continuous monitoring and evaluation should be conducted to ensure reliability and fairness.

How to Get Started with the Model

Use the code snippet below to get started with the model:

from transformers import pipeline


model_name = "tferhan/finetuned_camb_intents"

nlp_pipeline = pipeline("text-classification", model_name)

questions = ["qu'est ce que open data", "je veux les informations de l'eau potable"]
results = nlp_pipeline_class(questions)

for result in results:
    print(result)

#{'label': 'LABEL_0', 'score': 0.9999700784683228} === general
#{'label': 'LABEL_1', 'score': 0.9994990825653076} === request_data

Training Details

Training Data

• Data Source: Generated using GPT-4o-mini with help from words and data from data.gov.ma.

Training Procedure

• Preprocessing:
  • Standard text preprocessing steps - tokenization, text cleaning, and normalization.
• Training Hyperparameters:
  • Epochs: 10
  • Train Batch Size: 4
  • Eval Batch Size: 4
  • Learning Rate: 2e-5
  • Evaluation Strategy: epoch
  • Weight Decay: 0.01
• Log History: log_history.json

Evaluation

Testing Data & Metrics

• Testing Data: Subset of the generated data based on data.gov.ma.
• Evaluation Metrics: Accuracy.

Results

• Maximum Accuracy: 94%