Upload generate_dataset.py

generate_dataset.py

@@ -0,0 +1,104 @@
+import pandas as pd
+import json
+import random
+import string
+import re
+
+# Configuration
+TARGET_TOTAL = 15000          # target samples per label
+MAX_PREFIXES = 5             # max negative prefixes per utterance
+CONTEXT_TURNS = 4            # max number of previous turns to include (0..CONTEXT_TURNS)
+random.seed(42)
+
+arabic_punctuations = '''`÷×؛<>_()*&^%][ـ،/:"؟.,'{}~¦+|!”…“–ـ'''
+english_punctuations = string.punctuation
+all_punctuations = arabic_punctuations + english_punctuations
+
+def normalize_text(text):
+    """Normalize excessive spacing but keep punctuation."""
+    if not isinstance(text, str):
+        return ""
+    text = re.sub(r"\s+", " ", text)
+    return text.strip()
+
+def generate_prefixes(sentence):
+    """Generate multiple incomplete prefixes (label=0)."""
+    words = sentence.split()
+    prefixes = []
+    total_prefixes = min(len(words) - 1, MAX_PREFIXES)
+    for k in range(1, total_prefixes + 1):
+        prefix = " ".join(words[:k]).rstrip(all_punctuations).strip()
+        if prefix:
+            prefixes.append(prefix)
+    return prefixes
+
+def build_eou_dataset_with_sliding_context(input_csv, output_json):
+    print(f"Loading dataset: {input_csv}")
+    df = pd.read_csv(input_csv)
+    text_col = "GroundTruthText"
+    time_col = "SegmentStart" if "SegmentStart" in df.columns else df.columns[2]
+    df = df.sort_values(by=["FileName", time_col])
+
+    print("Generating EOU samples with sliding window context…")
+
+    all_pos = []
+    all_neg = []
+
+    for filename, group in df.groupby("FileName"):
+        utterances = group[text_col].dropna().tolist()
+        utterances = [normalize_text(u) for u in utterances if u.strip()]
+
+        for idx, utt in enumerate(utterances):
+            # --- Create multiple context versions ---
+            for c_len in range(0, CONTEXT_TURNS + 1):
+                context_start = max(0, idx - c_len)
+                context_turns = utterances[context_start:idx]
+
+                if context_turns:
+                    context_text = " [SEP] ".join(context_turns)
+                    full_text = f"{context_text} [SEP] {utt}"
+                else:
+                    full_text = utt
+
+                # Positive sample
+                all_pos.append({
+                    "text": full_text,
+                    "label": 1
+                })
+
+                # Negative samples (prefixes)
+                prefixes = generate_prefixes(utt)
+                for p in prefixes:
+                    if context_turns:
+                        neg_text = f"{context_text} [SEP] {p}"
+                    else:
+                        neg_text = p
+                    all_neg.append({
+                        "text": neg_text,
+                        "label": 0
+                    })
+
+    # Shuffle
+    random.shuffle(all_pos)
+    random.shuffle(all_neg)
+
+    # Balance dataset to TARGET_TOTAL per label if needed
+    pos_final = all_pos[:TARGET_TOTAL]
+    neg_final = all_neg[:TARGET_TOTAL]
+
+    final = pos_final + neg_final
+    random.shuffle(final)
+
+    print(f"Total positive (label=1): {len(pos_final)}")
+    print(f"Total negative (label=0): {len(neg_final)}")
+    print(f"Final dataset size:       {len(final)}")
+
+    # Save
+    with open(output_json, "w", encoding="utf-8") as f:
+        json.dump(final, f, ensure_ascii=False, indent=2)
+
+    print(f"Saved to {output_json}")
+    print("Done!")
+
+# Run the script
+build_eou_dataset_with_sliding_context("train.csv", "eou_sada_augmented.json")