koustubhr/bilstm-smartmcq-solver

BiLSTM — Smart MCQ Solver

BiLSTM trained from scratch (custom tokenizer + vocab, no pretrained embeddings) on a binary-expanded multiple-choice QA dataset. Each (prompt, option) pair is scored independently; the top-3 highest-scoring options are returned as the prediction.

Checkpoint contents (bilstm.pth, loaded with torch.load):

• model_state_dict
• vocab — token-to-id mapping used at training time
• config — vocab_size, embed_dim, hidden_dim, num_layers, dropout, max_len
• best_val_map3 — validation MAP@3 at the checkpointed epoch

Used by the Streamlit demo.