DANGDOCAO commited on
Commit
dab1047
·
verified ·
1 Parent(s): 56909b7

Upload 6 files

Browse files
.gitattributes ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ Data_totalQCAtriples30K/30ktrain.json filter=lfs diff=lfs merge=lfs -text
2
+ Datatest1k/testorgin1k.json filter=lfs diff=lfs merge=lfs -text
3
+ Datatrain29k/29kcorpustag.json filter=lfs diff=lfs merge=lfs -text
Data_totalQCAtriples30K/30ktrain.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a1989e727e0bb58732b6cd569c87fe2e474816d69064d0769f34cd307544d2fa
3
+ size 11786416
Datatest1k/testorgin1k.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a1989e727e0bb58732b6cd569c87fe2e474816d69064d0769f34cd307544d2fa
3
+ size 11786416
Datatrain29k/29kcorpustag.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a1989e727e0bb58732b6cd569c87fe2e474816d69064d0769f34cd307544d2fa
3
+ size 11786416
README.md ADDED
@@ -0,0 +1,239 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # HVU_QA_NLP
2
+
3
+ **HVU_QA_NLP** is a project dedicated to sharing datasets and tools for **Question Generation Processing (NLP)**, developed and maintained by the research team at **Hung Vuong University (HVU), Phu Tho, Vietnam**.
4
+ This project is supported by **Hung Vuong University, Phu Tho, Vietnam**, with the aim of advancing research and applications in low-resource language processing, particularly for the Vietnamese language.
5
+
6
+ ---
7
+
8
+ ## 📚 Overview
9
+
10
+ This repository enables you to:
11
+ 1. Fine-tune the [VietAI/vit5-base](https://huggingface.co/VietAI/vit5-base) model on your own QA dataset.
12
+ 2. Generate multiple, diverse questions given a user-provided text passage (context).
13
+
14
+ ---
15
+
16
+ ## 📁 Dataset Format
17
+
18
+ Your dataset must follow the **SQuAD v2.0** JSON structure:
19
+
20
+ ```json
21
+ {
22
+ "version": "v2.0",
23
+ "data": [
24
+ {
25
+ "title": "Đồ uống nào của Việt Nam từng lọt top ngon nhất thế giới?",
26
+ "paragraphs": [
27
+ {
28
+ "qas": [
29
+ {
30
+ "id": "q1_1",
31
+ "question": "Đồ uống nào của Việt Nam từng lọt top ngon nhất thế giới?",
32
+ "answers": [
33
+ {
34
+ "text": "Theo bản đánh giá tháng 2/2023 của Taste Atlas...",
35
+ "answer_start": "Theo bản đánh giá tháng 2/2023 của Taste Atlas..."
36
+ }
37
+ ],
38
+ "is_impossible": false
39
+ }
40
+ ]
41
+ }
42
+ ]
43
+ }
44
+ ]
45
+ }
46
+ ```
47
+
48
+ **Required fields:**
49
+ - `title` → used as context
50
+ - `question` → target question
51
+ - `answers[0].text` → seed answer for training
52
+ - `is_impossible` → filter for valid QAs
53
+
54
+ **File name:** `30ktrain.json` (UTF-8)
55
+
56
+ ---
57
+
58
+ ## 📁 Datasets
59
+
60
+ This repository provides datasets for **training** and **evaluating** Vietnamese question generation models.
61
+
62
+ ### 🔹 `DataTotalQCAtriples30k/`
63
+ - **`30ktrain.json`** → 30,000 QCA triples for training.
64
+
65
+ ### 🔹 `Datatest1k/`
66
+ - **`testorgin1k.json`** → 1,000 examples for manual & automatic evaluation.
67
+
68
+ ### 🔹 `Datatrain29k/`
69
+ - **`29kcorpustag.json`** → 29,000 preprocessed QCA triples.
70
+
71
+ > All files are UTF-8 encoded and ready for direct use in NLP pipelines.
72
+
73
+ ---
74
+
75
+ ## 📊 Evaluation Results
76
+
77
+ We performed **manual evaluation on 500 samples** and **automatic evaluation on 1,000 samples**.
78
+
79
+ | Evaluation Type | Precision | Recall | F1-Score |
80
+ |------------------|-----------|--------|----------|
81
+ | Automatic (1000) | 0.85 | 0.83 | 0.84 |
82
+ | Manual (500) | 0.88 | 0.86 | 0.87 |
83
+
84
+ ---
85
+
86
+ ## 🔧 Installation
87
+
88
+ Create a virtual environment and install dependencies:
89
+
90
+ ### Windows (PowerShell)
91
+ ```powershell
92
+ python -m venv .venv
93
+ .venv\Scripts\Activate.ps1
94
+ python -m pip install --upgrade pip
95
+ pip install torch transformers datasets scikit-learn
96
+ ```
97
+
98
+ ### Linux / macOS
99
+ ```bash
100
+ python3 -m venv .venv
101
+ source .venv/bin/activate
102
+ python -m pip install --upgrade pip
103
+ pip install torch transformers datasets scikit-learn
104
+ ```
105
+
106
+ ---
107
+
108
+ ## 🚀 Training (Fine-tuning)
109
+
110
+ Make sure `30ktrain.json` is in the same folder as `fine_tune_qg.py`.
111
+
112
+ Run:
113
+ ```bash
114
+ python fine_tune_qg.py
115
+ ```
116
+
117
+ **What happens:**
118
+ - Loads `30ktrain.json`
119
+ - Splits into 80% train / 20% validation
120
+ - Tokenizes using `T5Tokenizer`
121
+ - Fine-tunes `VietAI/vit5-base` for 3 epochs
122
+ - Saves model + tokenizer to `t5-viet-qg-finetuned/`
123
+
124
+ **Key training parameters** (edit in `fine_tune_qg.py` if needed):
125
+ - `per_device_train_batch_size`: 1
126
+ - `learning_rate`: 2e-4
127
+ - `num_train_epochs`: 3
128
+ - `max_input_length`: 512
129
+ - `max_target_length`: 64
130
+
131
+ ---
132
+
133
+ ## 💡 Generating Questions
134
+
135
+ Once training is done, use `generate_question.py` to generate new questions.
136
+
137
+ Ensure:
138
+ - `MODEL_DIR` → `t5-viet-qg-finetuned/`
139
+ - `DATA_PATH` → `30ktrain.json`
140
+
141
+ Run:
142
+ ```bash
143
+ python generate_question.py
144
+ ```
145
+
146
+ Steps:
147
+ 1. Enter a context passage (Vietnamese)
148
+ 2. Enter number of questions (default 20, max 200)
149
+ 3. Script will:
150
+ - Find best match in dataset by title similarity
151
+ - Use matched answer + your context
152
+ - Generate multiple unique questions with top-k & top-p sampling
153
+ 4. Output lists generated questions.
154
+
155
+ ---
156
+
157
+ ## ⚙️ Generation Settings
158
+
159
+ In `generate_question.py`, tweak:
160
+ - `top_k` (default 60)
161
+ - `top_p` (default 0.95)
162
+ - `temperature` (default 0.9)
163
+ - `no_repeat_ngram_size` (default 3)
164
+ - `repetition_penalty` (default 1.12)
165
+
166
+ ---
167
+
168
+ ## 📂 Project Structure
169
+
170
+ ```
171
+ .
172
+ ├── fine_tune_qg.py
173
+ ├── generate_question.py
174
+ ├── 30ktrain.json
175
+ ├── t5-viet-qg-finetuned/
176
+ ├── README.md
177
+ └── LICENSE
178
+ ```
179
+
180
+ ---
181
+
182
+ ## 🔍 Example Usage
183
+
184
+ **Training**
185
+ ```bash
186
+ python fine_tune_qg.py
187
+ ```
188
+
189
+ **Generating**
190
+ ```bash
191
+ python generate_question.py
192
+ ```
193
+
194
+ Example:
195
+ ```
196
+ Nhập đoạn văn bản:
197
+ Cà phê sữa đá là đồ uống nổi tiếng ở Việt Nam.
198
+
199
+ Nhập vào số lượng câu hỏi bạn cần: 5
200
+
201
+ ✅ Các câu hỏi mới được sinh ra:
202
+ 1. Loại cà phê nào nổi tiếng ở Việt Nam?
203
+ 2. Tại sao cà phê sữa đá được yêu thích?
204
+ 3. Cà phê sữa đá gồm những nguyên liệu gì?
205
+ 4. Nguồn gốc của cà phê sữa đá là từ đâu?
206
+ 5. Cà phê sữa đá Việt Nam được pha chế như thế nào?
207
+ ```
208
+
209
+ ---
210
+
211
+ ## 🤝 Contribution
212
+
213
+ You’re welcome to:
214
+ - Open issues
215
+ - Submit pull requests
216
+ - Suggest new datasets
217
+
218
+ ---
219
+
220
+ ## 📄 License
221
+
222
+ Licensed under the MIT License – see the `LICENSE` file for details.
223
+
224
+ ---
225
+
226
+ ## 📬 Contact
227
+
228
+ - **Ha Nguyen-Tien** (Corresponding author)
229
+ Email: nguyentienha@hvu.edu.vn
230
+
231
+ - **Phuc Le-Hong**
232
+ Email: Lehongphuc20021408@gmail.com
233
+
234
+ - **DANG DO CAO**
235
+ Email: docaodang532001@gmail.com
236
+
237
+ ---
238
+
239
+ *This repository is part of an effort to advance Vietnamese NLP by making question generation more accessible for researchers and developers.*
fine_tune_qg.py ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import json
2
+ from datasets import Dataset
3
+ from sklearn.model_selection import train_test_split
4
+ from transformers import (
5
+ T5Tokenizer,
6
+ T5ForConditionalGeneration,
7
+ TrainingArguments,
8
+ Trainer
9
+ )
10
+
11
+ def load_squad_data(file_path):
12
+ with open(file_path, "r", encoding="utf-8") as f:
13
+ squad_data = json.load(f)
14
+
15
+ data = []
16
+ for article in squad_data["data"]:
17
+ context = article.get("title", "")
18
+ for paragraph in article["paragraphs"]:
19
+ for qa in paragraph["qas"]:
20
+ if not qa.get("is_impossible", False) and qa.get("answers"):
21
+ answer = qa["answers"][0]["text"]
22
+ question = qa["question"]
23
+ input_text = f"answer: {answer} context: {context}"
24
+ data.append({"input": input_text, "target": question})
25
+ return data
26
+
27
+ def preprocess_function(example, tokenizer, max_input_length=512, max_target_length=64):
28
+ model_inputs = tokenizer(
29
+ example["input"],
30
+ max_length=max_input_length,
31
+ padding="max_length",
32
+ truncation=True,
33
+ )
34
+ labels = tokenizer(
35
+ text_target=example["target"],
36
+ max_length=max_target_length,
37
+ padding="max_length",
38
+ truncation=True,
39
+ )
40
+ model_inputs["labels"] = labels["input_ids"]
41
+ return model_inputs
42
+
43
+ def main():
44
+ data_path = "30ktrain.json"
45
+ output_dir = "t5-viet-qg-finetuned"
46
+ logs_dir = "logs"
47
+ model_name = "VietAI/vit5-base"
48
+
49
+ print("📥 Tải mô hình và tokenizer...")
50
+ tokenizer = T5Tokenizer.from_pretrained(model_name)
51
+ model = T5ForConditionalGeneration.from_pretrained(model_name)
52
+
53
+ print("📚 Đọc và chia dữ liệu...")
54
+ raw_data = load_squad_data(data_path)
55
+ train_data, val_data = train_test_split(raw_data, test_size=0.2, random_state=42)
56
+
57
+ train_dataset = Dataset.from_list(train_data)
58
+ val_dataset = Dataset.from_list(val_data)
59
+
60
+ tokenized_train = train_dataset.map(
61
+ lambda x: preprocess_function(x, tokenizer),
62
+ batched=True,
63
+ remove_columns=["input", "target"]
64
+ )
65
+ tokenized_val = val_dataset.map(
66
+ lambda x: preprocess_function(x, tokenizer),
67
+ batched=True,
68
+ remove_columns=["input", "target"]
69
+ )
70
+
71
+ print("⚙️ Cấu hình huấn luyện...")
72
+ training_args = TrainingArguments(
73
+ output_dir=output_dir,
74
+ overwrite_output_dir=True,
75
+ per_device_train_batch_size=1,
76
+ gradient_accumulation_steps=1,
77
+ num_train_epochs=3,
78
+ learning_rate=2e-4,
79
+ weight_decay=0.01,
80
+ warmup_steps=0,
81
+ logging_dir=logs_dir,
82
+ logging_steps=10,
83
+ fp16=False
84
+ )
85
+
86
+ print("🚀 Huấn luyện mô hình...")
87
+ trainer = Trainer(
88
+ model=model,
89
+ args=training_args,
90
+ train_dataset=tokenized_train,
91
+ eval_dataset=tokenized_val, # vẫn truyền để bạn có thể eval thủ công sau
92
+ tokenizer=tokenizer,
93
+ )
94
+ trainer.train()
95
+
96
+ print("💾 Lưu mô hình...")
97
+ model.save_pretrained(output_dir)
98
+ tokenizer.save_pretrained(output_dir)
99
+ print("✅ Huấn luyện hoàn tất!")
100
+
101
+ if __name__ == "__main__":
102
+ main()
generate_question.py ADDED
@@ -0,0 +1,154 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # generate_question.py
2
+ import json
3
+ from difflib import SequenceMatcher
4
+ from transformers import T5Tokenizer, T5ForConditionalGeneration
5
+ from transformers.utils import logging as hf_logging
6
+
7
+ # ===== Giảm ồn cảnh báo của transformers =====
8
+ hf_logging.set_verbosity_error()
9
+
10
+ # ===== Cấu hình đường dẫn =====
11
+ MODEL_DIR = "t5-viet-qg-finetuned" # thư mục model đã fine-tune
12
+ DATA_PATH = "30ktrain.json" # file JSON dữ liệu gốc
13
+
14
+ # ===== Load model & tokenizer 1 lần =====
15
+ tokenizer = T5Tokenizer.from_pretrained(MODEL_DIR)
16
+ model = T5ForConditionalGeneration.from_pretrained(MODEL_DIR)
17
+
18
+ def find_best_match_from_context(user_context, squad_data):
19
+ """
20
+ Tìm bản ghi gần nhất dựa trên article.title (giữ đúng logic code gốc).
21
+ Trả về tuple (context_title, answer_text, question_text) hoặc None.
22
+ """
23
+ best_score, best_entry = 0.0, None
24
+ ui = user_context.lower()
25
+
26
+ for article in squad_data.get("data", []):
27
+ context_title = article.get("title", "")
28
+ score_title = SequenceMatcher(None, ui, context_title.lower()).ratio()
29
+
30
+ for paragraph in article.get("paragraphs", []):
31
+ for qa in paragraph.get("qas", []):
32
+ answers = qa.get("answers", [])
33
+ if not answers:
34
+ continue
35
+ answer_text = answers[0].get("text", "").strip()
36
+ question_text = qa.get("question", "").strip()
37
+
38
+ score = score_title
39
+ if score > best_score:
40
+ best_score = score
41
+ best_entry = (context_title, answer_text, question_text)
42
+
43
+ return best_entry
44
+
45
+ def _near_duplicate(q, seen, thr=0.90):
46
+ """Loại câu gần trùng dựa trên tỉ lệ giống nhau."""
47
+ for s in seen:
48
+ if SequenceMatcher(None, q, s).ratio() >= thr:
49
+ return True
50
+ return False
51
+
52
+ def generate_questions(user_context,
53
+ total_questions=20,
54
+ batch_size=10,
55
+ top_k=60,
56
+ top_p=0.95,
57
+ temperature=0.9,
58
+ max_input_len=512,
59
+ max_new_tokens=64):
60
+ # Load dữ liệu JSON
61
+ with open(DATA_PATH, "r", encoding="utf-8") as f:
62
+ squad_data = json.load(f)
63
+
64
+ # Tìm bản ghi gần giống nhất (KHÔNG in ra màn hình)
65
+ best_entry = find_best_match_from_context(user_context, squad_data)
66
+ if best_entry is None:
67
+ print("❌ Không tìm thấy dữ liệu phù hợp trong file JSON.")
68
+ return
69
+
70
+ _, answer, _ = best_entry # chỉ cần answer gốc để bảo toàn tính toàn vẹn
71
+
72
+ # Chuẩn bị input cho mô hình (đặt max_length để tránh cảnh báo truncate)
73
+ input_text = f"answer: {answer} context: {user_context}"
74
+ inputs = tokenizer(
75
+ input_text,
76
+ return_tensors="pt",
77
+ truncation=True,
78
+ max_length=max_input_len
79
+ )
80
+
81
+ # Sinh nhiều câu hỏi theo lô để tiết kiệm RAM
82
+ unique_questions = []
83
+ remaining = total_questions
84
+
85
+ while remaining > 0:
86
+ n = min(batch_size, remaining)
87
+ outputs = model.generate(
88
+ **inputs,
89
+ do_sample=True,
90
+ top_k=top_k,
91
+ top_p=top_p,
92
+ temperature=temperature,
93
+ max_new_tokens=max_new_tokens, # chiều dài phần sinh
94
+ num_return_sequences=n,
95
+ no_repeat_ngram_size=3,
96
+ repetition_penalty=1.12
97
+ )
98
+
99
+ for out in outputs:
100
+ q = tokenizer.decode(out, skip_special_tokens=True).strip()
101
+ if len(q) < 5: # lọc câu quá ngắn
102
+ continue
103
+ if not _near_duplicate(q, unique_questions, thr=0.90):
104
+ unique_questions.append(q)
105
+
106
+ remaining = total_questions - len(unique_questions)
107
+ if remaining <= 0:
108
+ break
109
+
110
+ # Cắt đúng số lượng yêu cầu
111
+ unique_questions = unique_questions[:total_questions]
112
+
113
+ print("✅ Các câu hỏi mới được sinh ra:")
114
+ for i, q in enumerate(unique_questions, 1):
115
+ print(f"{i}. {q}")
116
+
117
+ if __name__ == "__main__":
118
+ # Nhập đoạn văn
119
+ user_context = input("\nNhập đoạn văn bản:\n ").strip()
120
+
121
+ # Nhập số lượng câu hỏi cần sinh (tùy biến)
122
+ raw_n = input("\nNhập vào số lượng câu hỏi bạn cần:").strip()
123
+ if raw_n == "":
124
+ total_questions = 20
125
+ else:
126
+ try:
127
+ total_questions = int(raw_n)
128
+ except ValueError:
129
+ print("⚠️ Giá trị không hợp lệ. Dùng mặc định 20.")
130
+ total_questions = 20
131
+
132
+ # Ràng buộc an toàn
133
+ if total_questions < 1:
134
+ total_questions = 1
135
+ if total_questions > 200:
136
+ print("⚠️ Giới hạn tối đa 200 câu. Sẽ sinh 200 câu.")
137
+ total_questions = 200
138
+
139
+ # batch_size tự động theo quy mô
140
+ batch_size = 10 if total_questions >= 30 else min(10, total_questions)
141
+
142
+ # Thông báo phân tích
143
+ print("\n🔍 Đang phân tích dữ liệu...\n")
144
+
145
+ generate_questions(
146
+ user_context=user_context,
147
+ total_questions=total_questions,
148
+ batch_size=batch_size,
149
+ top_k=60,
150
+ top_p=0.95,
151
+ temperature=0.9,
152
+ max_input_len=512,
153
+ max_new_tokens=64
154
+ )