import spaces import torch import gradio as gr from transformers import ( AutoModelForQuestionAnswering, AutoTokenizer ) # -------------------------------------------------- # 1. Load the RoBERTa SQuAD2 model # -------------------------------------------------- mdl_name = "deepset/roberta-base-squad2" device = torch.device( "cuda" if torch.cuda.is_available() else "cpu" ) print("Loading model:", mdl_name) print("Using device:", device) tokenizer = AutoTokenizer.from_pretrained(mdl_name) model = AutoModelForQuestionAnswering.from_pretrained( mdl_name ) model.to(device) model.eval() print("Model loaded successfully!") # -------------------------------------------------- # 2. Question answering function # -------------------------------------------------- def answer_question(question, context): if not question or not context: return "Please enter both a question and context." # Tokenize and split long contexts into overlapping chunks encoded = tokenizer( question, context, max_length=512, truncation="only_second", stride=128, return_overflowing_tokens=True, return_offsets_mapping=True, padding="max_length", return_tensors="pt" ) # Save character offsets and sequence information offsets = encoded.pop("offset_mapping").tolist() sequence_ids = [ encoded.sequence_ids(i) for i in range(encoded["input_ids"].shape[0]) ] # Remove metadata not accepted by the model encoded.pop("overflow_to_sample_mapping", None) # Move model inputs to CPU/GPU model_inputs = { key: value.to(device) for key, value in encoded.items() } # Run model inference with torch.inference_mode(): outputs = model(**model_inputs) best_difference = float("-inf") best_answer = "" # Examine every context chunk for i in range(len(sequence_ids)): start_logits = outputs.start_logits[i] end_logits = outputs.end_logits[i] input_ids = encoded["input_ids"][i] # Find CLS token for the no-answer score cls_index = ( input_ids == tokenizer.cls_token_id ).nonzero(as_tuple=True)[0][0].item() null_score = ( start_logits[cls_index].item() + end_logits[cls_index].item() ) valid_positions = [ j for j, seq_id in enumerate(sequence_ids[i]) if seq_id == 1 and offsets[i][j][1] > offsets[i][j][0] ] chunk_best_score = float("-inf") chunk_best_answer = "" # Find the best answer span for start in valid_positions: for end in valid_positions: if end < start: continue if end - start + 1 > 30: break score = ( start_logits[start].item() + end_logits[end].item() ) if score > chunk_best_score: start_char = offsets[i][start][0] end_char = offsets[i][end][1] chunk_best_score = score chunk_best_answer = context[ start_char:end_char ] # Compare answer score against no-answer score difference = chunk_best_score - null_score if difference > best_difference: best_difference = difference best_answer = chunk_best_answer # SQuAD2 supports questions without an answer if best_difference <= 0 or not best_answer.strip(): return "No answer found in the provided context." return best_answer.strip() # -------------------------------------------------- # 3. Gradio user interface # -------------------------------------------------- demo = gr.Interface( fn=answer_question, inputs=[ gr.Textbox( label="Question", placeholder="Enter your question..." ), gr.Textbox( label="Context", lines=10, placeholder="Enter the information to search..." ) ], outputs=gr.Textbox(label="Answer"), title="RoBERTa Question Answering", description="Ask questions about the supplied context using deepset/roberta-base-squad2." ) # -------------------------------------------------- # 4. Launch application # -------------------------------------------------- if __name__ == "__main__": demo.launch()