scandukuri/user-behavior-classifier-Llama-3.1-8B-Instruct-scalar-head-attention

This model is a classifier for user queries and followups built on meta-llama/Llama-3.1-8B-Instruct. After expanding the source datasets lmsys/lmsys-chat-1m and Magpie-Align/Magpie-Air-MT-300K-v0.1 with each user turn labeled as ground truth 'real' and 'synthetic' queries respectively, a scalar head was inserted on top of the instruction-tuned model and trained with nn.BCEWithLogitsLoss (a sigmoid activation and binary cross-entropy loss) to distinguish future 'real' and 'synthetic' queries.

In this model ablation, all attention modules were unfrozen during training as well.