{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"gpu","dataSources":[{"sourceType":"datasetVersion","sourceId":5807888,"datasetId":3335974,"databundleVersionId":5884620},{"sourceType":"datasetVersion","sourceId":11282395,"datasetId":7053929,"databundleVersionId":11701009},{"sourceType":"datasetVersion","sourceId":11282071,"datasetId":7053688,"databundleVersionId":11700642},{"sourceType":"datasetVersion","sourceId":11280450,"datasetId":7052486,"databundleVersionId":11698796}],"dockerImageVersionId":30919,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"import pandas as pd\n\ndf = pd.read_csv('/kaggle/input/d-qa-v3/distortion_processed.csv')\n\ndf['Question']=\"What is the distorted part?\"\n\nimport re\n\ndef find_answer_start(context, answer):\n    match = re.search(re.escape(answer.strip()), context)\n    return match.start() if match else -1\n\n# Chuyển đổi DataFrame thành định dạng cho mô hình QA\nquestions = df[\"Question\"].tolist()\ncontexts = df[\"Patient Question\"].tolist()\nanswer_ = df[\"processed_substrings\"].tolist()\n\nanswers = []\nfor i, row in df.iterrows():\n    if pd.isna(row[\"processed_substrings\"]):  # Không có câu trả lời\n        answer = {\"text\": [\"\"], \"answer_start\": [-1]}  \n    else:\n        start_idx = find_answer_start(row[\"Patient Question\"], answer_[i])\n        answer = {\n            \"text\": [row[\"processed_substrings\"]],\n            \"answer_start\": [start_idx]\n        }\n    answers.append(answer)\n\n\n\n# Tạo dataset theo định dạng của Hugging Face\nqa_data = {\n    \"distortion\": df[\"Dominant Distortion\"].tolist(),\n    \"question\": questions,\n    \"context\": df[\"Patient Question\"].tolist(),\n    \"answers\": answers\n}\n\n#Chuyển sang Dataset của Hugging Face\nfrom datasets import Dataset\ndataset = Dataset.from_dict(qa_data)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-09T02:49:04.123373Z","iopub.execute_input":"2026-04-09T02:49:04.123801Z","iopub.status.idle":"2026-04-09T02:49:07.702176Z","shell.execute_reply.started":"2026-04-09T02:49:04.123770Z","shell.execute_reply":"2026-04-09T02:49:07.701266Z"}},"outputs":[],"execution_count":1},{"cell_type":"code","source":"# dataset = dataset.train_test_split(test_size=0.2, seed=42)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-09T02:49:07.703325Z","iopub.execute_input":"2026-04-09T02:49:07.703649Z","iopub.status.idle":"2026-04-09T02:49:07.707440Z","shell.execute_reply.started":"2026-04-09T02:49:07.703626Z","shell.execute_reply":"2026-04-09T02:49:07.706400Z"}},"outputs":[],"execution_count":2},{"cell_type":"code","source":"# dataset","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-09T02:49:07.709576Z","iopub.execute_input":"2026-04-09T02:49:07.709829Z","iopub.status.idle":"2026-04-09T02:49:07.735105Z","shell.execute_reply.started":"2026-04-09T02:49:07.709809Z","shell.execute_reply":"2026-04-09T02:49:07.734071Z"}},"outputs":[],"execution_count":3},{"cell_type":"code","source":"def predict_answer(question, context):\n    # Tokenize with overflow & offset mapping\n    inputs = tokenizer(\n        question,\n        context,\n        return_tensors=\"pt\",\n        max_length=512,\n        truncation=\"only_second\",\n        stride=256,\n        return_overflowing_tokens=True,\n        return_offsets_mapping=True,\n        padding=\"max_length\",\n    )\n\n    # Pop những thông tin không phải input cho model\n    offset_mapping = inputs.pop(\"offset_mapping\")\n    overflow_mapping = inputs.pop(\"overflow_to_sample_mapping\")\n\n    input_ids = inputs[\"input_ids\"]\n    # print(\"Số đoạn context:\", input_ids.shape[0])\n    inputs = {k: v.to(device) for k, v in inputs.items()}\n\n    with torch.no_grad():\n        outputs = model(**inputs)\n\n    start_logits = outputs.start_logits\n    end_logits = outputs.end_logits\n\n    best_score = float('-inf')\n    best_answer = \"\"\n\n    for i in range(len(start_logits)):\n        start_logit = start_logits[i]\n        end_logit = end_logits[i]\n        offsets = offset_mapping[i]\n\n        start_index = torch.argmax(start_logit).item()\n        end_index = torch.argmax(end_logit).item()\n\n        # Kiểm tra chỉ số có hợp lệ không\n        if (\n            start_index >= len(offsets)\n            or end_index >= len(offsets)\n            or offsets[start_index] is None\n            or offsets[end_index] is None\n        ):\n            continue\n\n        start_char = offsets[start_index][0]\n        end_char = offsets[end_index][1]\n\n        score = start_logit[start_index] + end_logit[end_index]\n        # print(score)\n        if score > best_score and start_char < end_char:\n            best_score = score\n            best_answer = context[start_char:end_char]\n\n    return best_answer","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-09T02:49:07.736360Z","iopub.execute_input":"2026-04-09T02:49:07.736718Z","iopub.status.idle":"2026-04-09T02:49:07.762185Z","shell.execute_reply.started":"2026-04-09T02:49:07.736682Z","shell.execute_reply":"2026-04-09T02:49:07.761018Z"}},"outputs":[],"execution_count":4},{"cell_type":"code","source":"from collections import Counter\nimport string\n\n# Hàm tính toán Exact Match và F1 (như đã trình bày trước đó)\ndef calculate_exact_match(predictions, ground_truths):\n    exact_match = 0\n    for pred, truth in zip(predictions, ground_truths):\n        if pred.strip().lower() == truth.strip().lower():\n            exact_match += 1\n    return exact_match / len(predictions)\n\ndef calculate_f1_score(predictions, ground_truths):\n    def f1(pred, truth):\n        pred_tokens = normalize_text(pred).split()\n        truth_tokens = normalize_text(truth).split()\n\n        # Trường hợp cả hai chuỗi đều rỗng\n        if not pred_tokens and not truth_tokens:\n            return 1.0  # Có thể cho là F1 score hoàn hảo trong trường hợp này\n\n        common_tokens = Counter(pred_tokens) & Counter(truth_tokens)\n        num_common = sum(common_tokens.values())\n\n        if num_common == 0:\n            return 0.0  # Trả về giá trị 0 nếu không có sự giao nhau\n\n        precision = num_common / len(pred_tokens) if len(pred_tokens) > 0 else 0\n        recall = num_common / len(truth_tokens) if len(truth_tokens) > 0 else 0\n\n        if precision + recall == 0:\n            return 0.0  # Tránh trường hợp chia cho 0\n\n        return 2 * (precision * recall) / (precision + recall)\n\n    f1_scores = [f1(pred, truth) for pred, truth in zip(predictions, ground_truths)]\n    return sum(f1_scores) / len(f1_scores)\n\n\ndef normalize_text(text):\n    return text.translate(str.maketrans(\"\", \"\", string.punctuation)).lower()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-09T02:49:07.762952Z","iopub.execute_input":"2026-04-09T02:49:07.763265Z","iopub.status.idle":"2026-04-09T02:49:07.821812Z","shell.execute_reply.started":"2026-04-09T02:49:07.763226Z","shell.execute_reply":"2026-04-09T02:49:07.820355Z"}},"outputs":[],"execution_count":5},{"cell_type":"code","source":"import torch\nimport numpy as np\nfrom datasets import Dataset\nfrom transformers import AutoTokenizer, AutoModelForQuestionAnswering\nfrom torch.utils.data import DataLoader\nfrom transformers import default_data_collator\nfrom tqdm.auto import tqdm\n\n# Tokenizer\nmodel_checkpoint = \"bert-base-cased\"\ntokenizer = AutoTokenizer.from_pretrained(model_checkpoint)\n\n# Preprocessing function\ndef preprocess_examples(examples):\n    questions = [q.strip() for q in examples[\"question\"]]\n    inputs = tokenizer(\n        questions,\n        examples[\"context\"],\n        max_length=512,\n        truncation=\"only_second\",\n        stride=256,\n        return_overflowing_tokens=True,\n        return_offsets_mapping=True,\n        padding=\"max_length\",\n    )\n    \n    offset_mapping = inputs.pop(\"offset_mapping\")\n    sample_map = inputs.pop(\"overflow_to_sample_mapping\")\n    answers = examples[\"answers\"]\n    start_positions, end_positions = [], []\n    \n    for i, offsets in enumerate(offset_mapping):\n        sample_idx = sample_map[i]\n        answer = answers[sample_idx]\n        if len(answer[\"text\"]) == 0:\n            start_positions.append(0)\n            end_positions.append(0)\n        else:\n            start_char = answer[\"answer_start\"][0]\n            end_char = start_char + len(answer[\"text\"][0])\n            sequence_ids = inputs.sequence_ids(i)\n            \n            idx = 0\n            while sequence_ids[idx] != 1:\n                idx += 1\n            context_start = idx\n            while idx < len(sequence_ids) and sequence_ids[idx] == 1:\n                idx += 1\n            context_end = idx - 1\n            \n            if offsets[context_start][0] > start_char or offsets[context_end][1] < end_char:\n                start_positions.append(0)\n                end_positions.append(0)\n            else:\n                idx = context_start\n                while idx <= context_end and offsets[idx][0] <= start_char:\n                    idx += 1\n                start_positions.append(idx - 1)\n                \n                idx = context_end\n                while idx >= context_start and offsets[idx][1] >= end_char:\n                    idx -= 1\n                end_positions.append(idx + 1)\n    \n    inputs[\"start_positions\"] = start_positions\n    inputs[\"end_positions\"] = end_positions\n    return inputs\n\n# Apply preprocessing\ndataset_processed = dataset.map(preprocess_examples, batched=True, remove_columns=dataset.column_names)\ndataset_splited = dataset_processed.train_test_split(test_size=0.2, seed=42)\n\ndataset.set_format(\"torch\")\ntrain_dataloader = DataLoader(dataset_splited['train'], batch_size=3, shuffle=True, collate_fn=default_data_collator)\ntest_dataloader = DataLoader(dataset_splited['test'], batch_size=3, shuffle=False, collate_fn=default_data_collator)\n\n# Model and optimizer\nmodel = AutoModelForQuestionAnswering.from_pretrained(model_checkpoint)\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\nmodel.to(device)\noptimizer = torch.optim.AdamW(model.parameters(), lr=2e-5)\n\n# Training loop\nnum_epochs = 20\nprogress_bar = tqdm(range(len(train_dataloader) * num_epochs))\n\ndataset = Dataset.from_dict(qa_data)\ndataset = dataset.train_test_split(test_size=0.2, seed=42)\n\nfor epoch in range(num_epochs):\n    model.train()\n    for batch in train_dataloader:\n        batch = {k: v.to(device) for k, v in batch.items()}\n        optimizer.zero_grad()\n        outputs = model(**batch)\n        loss = outputs.loss\n        loss.backward()\n        optimizer.step()\n        progress_bar.update(1)\n    predictions = []\n    references = []\n    model.eval()\n    # Lặp qua từng dòng trong tập test\n    for example in dataset['test']:\n        question = example[\"question\"]\n        context = example[\"context\"]\n        ground_truth = example[\"answers\"][\"text\"][0]\n    \n        # Dự đoán câu trả lời\n        result = predict_answer(question, context)\n    \n        # Thêm vào danh sách đánh giá\n        predictions.append(result)\n        references.append(ground_truth)\n    \n    \n    # Tính Exact Match và F1\n    em = calculate_exact_match(predictions, references)\n    f1 = calculate_f1_score(predictions, references)\n    \n    print(f\"Exact Match: {em * 100:.2f}%\")\n    print(f\"F1 Score: {f1:.4f}\")\n    print(f\"Epoch {epoch + 1} completed.\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-09T02:49:08.670359Z","iopub.execute_input":"2026-04-09T02:49:08.670795Z","execution_failed":"2026-04-08T19:52:18.847Z"}},"outputs":[{"output_type":"display_data","data":{"text/plain":"tokenizer_config.json:   0%|          | 0.00/49.0 [00:00<?, ?B/s]","application/vnd.jupyter.widget-view+json":{"version_major":2,"version_minor":0,"model_id":"4459cdddee0241a2b16fe972fa08c58e"}},"metadata":{}},{"output_type":"display_data","data":{"text/plain":"config.json:   0%|          | 0.00/570 [00:00<?, ?B/s]","application/vnd.jupyter.widget-view+json":{"version_major":2,"version_minor":0,"model_id":"3e9bcdf9144749fa8259f0eaad45fdad"}},"metadata":{}},{"output_type":"display_data","data":{"text/plain":"vocab.txt: 0.00B [00:00, ?B/s]","application/vnd.jupyter.widget-view+json":{"version_major":2,"version_minor":0,"model_id":"01fbb2df8b8a403dbfc30ffc2f5260f3"}},"metadata":{}},{"output_type":"display_data","data":{"text/plain":"tokenizer.json: 0.00B [00:00, ?B/s]","application/vnd.jupyter.widget-view+json":{"version_major":2,"version_minor":0,"model_id":"e857c1b04ebf46ceafb90f3190c7151d"}},"metadata":{}},{"output_type":"display_data","data":{"text/plain":"Map:   0%|          | 0/2530 [00:00<?, ? examples/s]","application/vnd.jupyter.widget-view+json":{"version_major":2,"version_minor":0,"model_id":"4a73070a62df42788f1144353d05b292"}},"metadata":{}},{"output_type":"display_data","data":{"text/plain":"model.safetensors:   0%|          | 0.00/436M [00:00<?, ?B/s]","application/vnd.jupyter.widget-view+json":{"version_major":2,"version_minor":0,"model_id":"fc08de0a23244b0ca9f6e3b1cdc2a7c7"}},"metadata":{}},{"name":"stderr","text":"Some weights of BertForQuestionAnswering were not initialized from the model checkpoint at bert-base-cased and are newly initialized: ['qa_outputs.bias', 'qa_outputs.weight']\nYou should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n","output_type":"stream"},{"output_type":"display_data","data":{"text/plain":"  0%|          | 0/13960 [00:00<?, ?it/s]","application/vnd.jupyter.widget-view+json":{"version_major":2,"version_minor":0,"model_id":"372a4f27c1104f01985e59672266f38d"}},"metadata":{}},{"name":"stdout","text":"Exact Match: 40.12%\nF1 Score: 0.4749\nEpoch 1 completed.\n","output_type":"stream"}],"execution_count":null},{"cell_type":"code","source":"# Save model\nmodel.save_pretrained(\"bert-finetuned-distortion\")\ntokenizer.save_pretrained(\"bert-finetuned-distortion\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-29T04:02:20.639851Z","iopub.status.idle":"2025-09-29T04:02:20.640119Z","shell.execute_reply":"2025-09-29T04:02:20.640009Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"\n# def predict_answer(question, context):\n#     inputs = tokenizer(question, context, return_tensors=\"pt\", truncation=True, max_length=384)\n#     inputs = {k: v.to(device) for k, v in inputs.items()}\n    \n#     with torch.no_grad():\n#         outputs = model(**inputs)\n    \n#     start_logits = outputs.start_logits\n#     end_logits = outputs.end_logits\n#     start_index = torch.argmax(start_logits, dim=-1).item()\n#     end_index = torch.argmax(end_logits, dim=-1).item()\n    \n#     all_tokens = tokenizer.convert_ids_to_tokens(inputs[\"input_ids\"].squeeze())\n#     answer = tokenizer.convert_tokens_to_string(all_tokens[start_index:end_index+1])\n    \n#     return answer\n\n# def predict_answer(question, context):\n#     # Tokenize with offset mapping to align tokens with original text\n#     inputs = tokenizer(\n#         question,\n#         context,\n#         return_tensors=\"pt\",\n#         max_length=512,\n#         truncation=\"only_second\",\n#         stride=256,\n#         return_offsets_mapping=True,\n#         padding=\"max_length\",\n#     )\n\n#     offset_mapping = inputs.pop(\"offset_mapping\")\n#     input_ids = inputs[\"input_ids\"]\n#     inputs = {k: v.to(device) for k, v in inputs.items()}\n\n#     with torch.no_grad():\n#         outputs = model(**inputs)\n\n#     start_logits = outputs.start_logits\n#     end_logits = outputs.end_logits\n\n#     # Get the most probable start and end token positions\n#     start_index = torch.argmax(start_logits, dim=-1).item()\n#     end_index = torch.argmax(end_logits, dim=-1).item()\n\n#     # Extract character positions from offset mapping\n#     offsets = offset_mapping[0]\n#     start_char = offsets[start_index][0]\n#     end_char = offsets[end_index][1]\n\n#     # Extract the predicted answer from the original context\n#     answer = context[start_char:end_char]\n\n#     return answer\n\n\n\n# Example inference\n# for i in range(15):\n#     question = qa_data[\"question\"][i]\n#     context = qa_data[\"context\"][i]\n#     print(context)\n#     print(qa_data[\"answers\"][i])\n#     predicted_answer = predict_answer(question, context)\n#     print(f\"Question: {question}\")\n#     print(f\"Predicted Answer: {predicted_answer}\\n\")\n    ","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-09T02:38:13.063799Z","iopub.execute_input":"2026-04-09T02:38:13.064239Z","iopub.status.idle":"2026-04-09T02:38:13.074009Z","shell.execute_reply.started":"2026-04-09T02:38:13.064205Z","shell.execute_reply":"2026-04-09T02:38:13.072689Z"}},"outputs":[],"execution_count":7},{"cell_type":"code","source":"# # !pip install evaluate\n# import evaluate\n\n# # Load metrics\n# squad_metric = evaluate.load(\"squad\")\n\n# # Chuẩn bị prediction và references\n# predictions = []\n# references = []\n\n# for i in range(len(qa_data[\"question\"])):\n#     question = qa_data[\"question\"][i]\n#     context = qa_data[\"context\"][i]\n#     ground_truth = qa_data[\"answers\"][i][\"text\"][0]\n\n#     result = predict_answer(question, context)\n\n#     # Thêm vào danh sách đánh giá\n#     predictions.append({\n#         \"id\": str(i),\n#         \"prediction_text\": result\n#     })\n#     references.append({\n#         \"id\": str(i),\n#         \"answers\": {\n#             \"answer_start\": [qa_data[\"answers\"][i][\"answer_start\"][0]],\n#             \"text\": [ground_truth]\n#         }\n#     })\n\n# # Tính EM và F1\n# results = squad_metric.compute(predictions=predictions, references=references)\n# print(f\"Exact Match: {results['exact_match']:.2f}\")\n# print(f\"F1 Score: {results['f1']:.2f}\")\nfrom collections import Counter\nimport string\n\n# Hàm tính toán Exact Match và F1 (như đã trình bày trước đó)\ndef calculate_exact_match(predictions, ground_truths):\n    exact_match = 0\n    for pred, truth in zip(predictions, ground_truths):\n        if pred.strip().lower() == truth.strip().lower():\n            exact_match += 1\n    return exact_match / len(predictions)\n\ndef calculate_f1_score(predictions, ground_truths):\n    def f1(pred, truth):\n        pred_tokens = normalize_text(pred).split()\n        truth_tokens = normalize_text(truth).split()\n\n        # Trường hợp cả hai chuỗi đều rỗng\n        if not pred_tokens and not truth_tokens:\n            return 1.0  # Có thể cho là F1 score hoàn hảo trong trường hợp này\n\n        common_tokens = Counter(pred_tokens) & Counter(truth_tokens)\n        num_common = sum(common_tokens.values())\n\n        if num_common == 0:\n            return 0.0  # Trả về giá trị 0 nếu không có sự giao nhau\n\n        precision = num_common / len(pred_tokens) if len(pred_tokens) > 0 else 0\n        recall = num_common / len(truth_tokens) if len(truth_tokens) > 0 else 0\n\n        if precision + recall == 0:\n            return 0.0  # Tránh trường hợp chia cho 0\n\n        return 2 * (precision * recall) / (precision + recall)\n\n    f1_scores = [f1(pred, truth) for pred, truth in zip(predictions, ground_truths)]\n    return sum(f1_scores) / len(f1_scores)\n\n\ndef normalize_text(text):\n    return text.translate(str.maketrans(\"\", \"\", string.punctuation)).lower()\n\n#Chuẩn bị dữ liệu từ qa_data\n# predictions = []\n# references = []\n\n# for i in range(len(qa_data[\"question\"])):\n#     question = qa_data[\"question\"][i]\n#     context = qa_data[\"context\"][i]\n#     ground_truth = qa_data[\"answers\"][i][\"text\"][0]\n\n#     # Dự đoán câu trả lời cho từng câu hỏi\n#     result = predict_answer(question, context)\n#     qa_data[\"predict\"][i]= result\n#     # Thêm vào danh sách đánh giá\n#     predictions.append(result)\n#     references.append(ground_truth)\n\n# predictions = []\n# references = []\n\n# # Lặp qua từng dòng trong tập test\n# for example in dataset['test']:\n#     question = example[\"question\"]\n#     context = example[\"context\"]\n#     ground_truth = example[\"answers\"][\"text\"][0]\n\n#     # Dự đoán câu trả lời\n#     result = predict_answer(question, context)\n\n#     # Thêm vào danh sách đánh giá\n#     predictions.append(result)\n#     references.append(ground_truth)\n\n\n# # Tính Exact Match và F1\n# em = calculate_exact_match(predictions, references)\n# f1 = calculate_f1_score(predictions, references)\n\n# print(f\"Exact Match: {em * 100:.2f}%\")\n# print(f\"F1 Score: {f1:.4f}\")\n\n# qa_data[\"predict\"] = []  # Tạo danh sách rỗng trước\n\n# for i in range(len(qa_data[\"question\"])):\n#     question = qa_data[\"question\"][i]\n#     context = qa_data[\"context\"][i]\n#     ground_truth = qa_data[\"answers\"][i][\"text\"][0]\n\n#     # Dự đoán câu trả lời cho từng câu hỏi\n#     result = predict_answer(question, context)\n\n#     qa_data[\"predict\"].append(result)  # Thêm kết quả dự đoán\n\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-29T04:49:23.854050Z","iopub.execute_input":"2025-09-29T04:49:23.854337Z","iopub.status.idle":"2025-09-29T04:49:23.862356Z","shell.execute_reply.started":"2025-09-29T04:49:23.854316Z","shell.execute_reply":"2025-09-29T04:49:23.861378Z"}},"outputs":[],"execution_count":4},{"cell_type":"code","source":"import pandas as pd\n\n# Chuyển dict thành DataFrame\ndf_qa = pd.DataFrame(qa_data)\n","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.991Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"df_qa.to_csv('distorted.csv', index = False)","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.991Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":" df['Distorted part']=df['Distorted part'].fillna('')","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.991Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"references","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.991Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"qa_data[\"context\"][91]","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.991Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"df.head()","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.992Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"df['Distorted part'][91]","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.992Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"df['Longest Match'][91]","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.992Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"qa_data[\"answers\"][91]","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.992Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"for i in range(len(references)):\n    print(references[i])\n    print(predictions[i])","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.992Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"predictions[2]","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.992Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"references","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.992Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"predictions[0]","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.992Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"references[0]","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.992Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Load pipeline\nquestion_answerer = pipeline(\"question-answering\", model=model, tokenizer=tokenizer)\n\n# Example inference\nfor i in range(len(qa_data[\"question\"])):\n    question = qa_data[\"question\"][i]\n    context = qa_data[\"context\"][i]\n    result = question_answerer(question=question, context=context)\n    print(\"Answer:\", result[\"answer\"])\n    print(\"Start index:\", result[\"start\"])\n    print(\"End index:\", result[\"end\"])\n    print(\"Confidence score:\", result[\"score\"])\n    print(f\"Question: {question}\")\n    print(f\"Predicted Answer: {result['answer']}\\n\")\n    break","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.992Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"len(\"I don’t really know how to explain the situation. \")","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.992Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"qa_data","metadata":{"trusted":true,"execution":{"execution_failed":"2025-05-19T01:36:39.993Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}