fine-tune & evaluation

fine_tune_and_evaluation.py

@@ -0,0 +1,129 @@
+from transformers import BartTokenizer, BartForConditionalGeneration, TrainingArguments, Trainer
+import pandas as pd
+from datasets import Dataset, Features, Value
+import evaluate
+import nltk
+import json
+import os
+import random
+
+nltk.download('punkt')
+
+# === CONFIGURATION ===
+train_file = r"C:/Users/aditi/OneDrive/Desktop/train_v0.2 QuaC.json"
+model_name = "voidful/bart-eqg-question-generator"
+output_dir = "./bart-eqg-finetuned-500"
+
+# === FILE CHECK ===
+if not os.path.exists(train_file):
+    raise FileNotFoundError(f"File not found at: {train_file}")
+
+# === LOAD DATA ===
+with open(train_file, 'r', encoding='utf-8') as f:
+    quac_data = json.load(f)
+
+# === EXTRACT 500 Q&A PAIRS ===
+data = []
+for item in quac_data.get("data", []):
+    for paragraph in item.get("paragraphs", []):
+        context = paragraph.get("context", "")
+        for qa in paragraph.get("qas", []):
+            question = qa.get("question", "")
+            answer = qa.get("answers", [{}])[0].get("text", "") if qa.get("answers") else ""
+            if context and question and answer:
+                data.append({"context": context, "question": question, "answer": answer})
+
+random.seed(42)
+random.shuffle(data)
+data = data[:500]
+
+# === CREATE DATASET ===
+df = pd.DataFrame(data)[["context", "question", "answer"]]
+features = Features({
+    "context": Value("string"),
+    "question": Value("string"),
+    "answer": Value("string")
+})
+dataset = Dataset.from_pandas(df, features=features)
+train_test_split = dataset.train_test_split(test_size=0.2, seed=42)
+train_dataset = train_test_split["train"]
+eval_dataset = train_test_split["test"]
+
+print(f"Train size: {len(train_dataset)} | Eval size: {len(eval_dataset)}")
+
+# === LOAD MODEL AND TOKENIZER ===
+try:
+    tokenizer = BartTokenizer.from_pretrained(model_name)
+    model = BartForConditionalGeneration.from_pretrained(model_name)
+except Exception as e:
+    raise RuntimeError(f"Could not load model or tokenizer: {e}")
+
+# === PREPROCESS FUNCTION ===
+def preprocess(example):
+    input_text = example['context']
+    target_text = example['question']
+    model_inputs = tokenizer(input_text, max_length=512, truncation=True, padding="max_length")
+    labels = tokenizer(target_text, max_length=64, truncation=True, padding="max_length")["input_ids"]
+    model_inputs["labels"] = labels
+    return model_inputs
+
+tokenized_train_dataset = train_dataset.map(preprocess, remove_columns=train_dataset.column_names, batched=True)
+tokenized_eval_dataset = eval_dataset.map(preprocess, remove_columns=eval_dataset.column_names, batched=True)
+
+# === METRIC COMPUTATION ===
+def compute_metrics(eval_pred):
+    preds, labels = eval_pred
+    decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
+    decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
+
+    bleu = evaluate.load("bleu")
+    rouge = evaluate.load("rouge")
+
+    bleu_score = bleu.compute(predictions=decoded_preds, references=decoded_labels)
+    rouge_score = rouge.compute(predictions=decoded_preds, references=decoded_labels)
+
+    return {
+        "bleu": bleu_score["bleu"],
+        "rouge1": rouge_score["rouge1"],
+        "rougeL": rouge_score["rougeL"]
+    }
+
+# === TRAINING ARGS === (no evaluation_strategy used)
+training_args = TrainingArguments(
+    output_dir=output_dir,
+    per_device_train_batch_size=2,
+    per_device_eval_batch_size=2,
+    num_train_epochs=3,
+    save_strategy="epoch",
+    save_total_limit=1,
+    logging_dir="./logs",
+    logging_steps=10,
+    fp16=False,
+    report_to="none"
+)
+
+# === TRAINER ===
+trainer = Trainer(
+    model=model,
+    args=training_args,
+    train_dataset=tokenized_train_dataset,
+    eval_dataset=tokenized_eval_dataset,
+    compute_metrics=compute_metrics
+)
+
+# === TRAIN & EVALUATE ===
+print("Fine-tuning started...")
+#trainer.train()
+trainer.train(resume_from_checkpoint=True)
+
+print("Running final evaluation...")
+results = trainer.evaluate()
+print("Final Evaluation Results:")
+for metric, score in results.items():
+    print(f"  {metric}: {score}")
+
+# === SAVE MODEL ===
+model.save_pretrained(os.path.join(output_dir, "final"))
+tokenizer.save_pretrained(os.path.join(output_dir, "final"))
+print("Fine-tuned model and tokenizer saved!")
+

requirements.txt

@@ -1,58 +1,26 @@
-altair==5.5.0
-attrs==25.3.0
-blinker==1.9.0
-cachetools==6.1.0
-certifi==2025.6.15
-charset-normalizer==3.4.2
-click==8.2.1
-colorama==0.4.6
-filelock==3.18.0
-fsspec==2025.5.1
-gitdb==4.0.12
-GitPython==3.1.44
-huggingface-hub==0.33.2
-idna==3.10
-Jinja2==3.1.6
-joblib==1.5.1
-jsonschema==4.24.0
-jsonschema-specifications==2025.4.1
-MarkupSafe==3.0.2
-mpmath==1.3.0
-narwhals==1.45.0
-networkx==3.5
+# Core libraries
+transformers==4.53.2
+datasets==4.0.0
+evaluate==0.4.5
 nltk==3.9.1
-numpy==2.3.1
-packaging==25.0
-pandas==2.3.0
-pillow==11.3.0
-protobuf==6.31.1
-pyarrow==20.0.0
-pydeck==0.9.1
-python-dateutil==2.9.0.post0
-pytz==2025.2
-PyYAML==6.0.2
-referencing==0.36.2
-regex==2024.11.6
-requests==2.32.4
-rpds-py==0.26.0
-safetensors==0.5.3
-scikit-learn==1.7.0
-scipy==1.16.0
-sentence-transformers==3.1.1
-setuptools==80.9.0
-six==1.17.0
-smmap==5.0.2
-streamlit==1.46.1
-sympy==1.14.0
-tenacity==9.1.2
-threadpoolctl==3.6.0
-tokenizers==0.15.2
-toml==0.10.2
-torch==2.7.1
-tornado==6.5.1
-tqdm==4.67.1
-transformers==4.39.3
-typing_extensions==4.14.0
-tzdata==2025.2
-urllib3==2.5.0
-watchdog==6.0.0
+pandas==2.3.1
+numpy>=1.17
+tqdm>=4.27
+scipy
+
+# PyTorch (CPU version)
+torch==2.3.0
+torchaudio==2.3.0
+torchvision==0.18.0
+
+# Hugging Face Hub
+huggingface-hub>=0.16.4
+safetensors>=0.4.3
+
+# Optional but useful
+pyarrow>=15.0.0
+regex
+filelock
+fsspec
+
+accelerate>=0.26.0