Bilic/NeuralChat-finetuned-for-fraud-detection

[image]

neural-chat-finetuned-bilic-v1

This model is a fine-tuned version of Intel/neural-chat-7b-v3-1 on our custom dataset.

Model description

This is a fine tuned version of the intel's Neuralchat model, specifically trained on a carefully curated dataset on fraud detection. We implemented a contextual based architecture to enable the model learn and be adept at understanding context within a conversation as opposed to the traditional rule based approach.

Intended uses & limitations

• —detecting fraudulent conversations in real-time
• —Giving a summary of conversations and suggestions
• —Understanding with high accuracy the context in a conversation to make better predictions

Training

50,000 synthetically conversations

Training hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 0.0002
• —trainbatchsize: 8
• —evalbatchsize: 8
• —seed: 42
• —optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• —lrschedulertype: cosine
• —training_steps: 250
• —mixedprecisiontraining: Native AMP

Framework versions

• —Transformers 4.36.0.dev0
• —Pytorch 2.1.0+cu118
• —Datasets 2.15.0
• —Tokenizers 0.15.0