diff --git a/bitnet_tools/multi_csv.py b/bitnet_tools/multi_csv.py
index ee48828..667813c 100644
--- a/bitnet_tools/multi_csv.py
+++ b/bitnet_tools/multi_csv.py
@@ -13,6 +13,8 @@
 from .analysis import _to_float
 
 CACHE_DIR = Path('.bitnet_cache')
+UNIQUE_BITMAP_SIZE = 65536
+TOP_VALUE_TRACK_CAP = 5000
 
 
 def _quantile(sorted_values: list[float], q: float) -> float:
@@ -110,6 +112,32 @@ def _infer_semantic_type(col: str, dtype: str, samples: list[str], unique_ratio:
     return 'text'
 
 
+def _update_unique_bitmap(bitmap: bytearray, value: str) -> None:
+    h = hashlib.sha1(value.encode('utf-8')).digest()
+    idx = int.from_bytes(h[:8], 'big') % UNIQUE_BITMAP_SIZE
+    bitmap[idx // 8] |= 1 << (idx % 8)
+
+
+def _estimate_unique_count(bitmap: bytearray) -> int:
+    set_bits = sum(bin(b).count('1') for b in bitmap)
+    if set_bits <= 0:
+        return 0
+    if set_bits >= UNIQUE_BITMAP_SIZE:
+        return UNIQUE_BITMAP_SIZE
+    zero_bits = UNIQUE_BITMAP_SIZE - set_bits
+    return max(1, int(round(-UNIQUE_BITMAP_SIZE * math.log(zero_bits / UNIQUE_BITMAP_SIZE))))
+
+
+def _update_bounded_counter(counter: Counter[str], value: str, other_holder: dict[str, int], cap: int) -> None:
+    if value in counter:
+        counter[value] += 1
+        return
+    if len(counter) < cap:
+        counter[value] += 1
+        return
+    other_holder['count'] += 1
+
+
 def _profile_csv_stream(
     path: Path,
     group_column: str | None = None,
@@ -125,8 +153,9 @@ def _profile_csv_stream(
 
         missing = {c: 0 for c in columns}
         non_missing = {c: 0 for c in columns}
-        unique_sets: dict[str, set[str]] = {c: set() for c in columns}
+        unique_bitmaps: dict[str, bytearray] = {c: bytearray(UNIQUE_BITMAP_SIZE // 8) for c in columns}
         value_counts: dict[str, Counter[str]] = {c: Counter() for c in columns}
+        value_overflow: dict[str, dict[str, int]] = {c: {'count': 0} for c in columns}
         value_samples: dict[str, list[str]] = {c: [] for c in columns}
 
         numeric_positive = {c: 0 for c in columns}
@@ -156,8 +185,8 @@ def _profile_csv_stream(
                     missing[col] += 1
                     continue
                 non_missing[col] += 1
-                unique_sets[col].add(raw)
-                value_counts[col][raw] += 1
+                _update_unique_bitmap(unique_bitmaps[col], raw)
+                _update_bounded_counter(value_counts[col], raw, value_overflow[col], TOP_VALUE_TRACK_CAP)
                 _reservoir_sample_str(value_samples[col], raw, non_missing[col], value_sample_cap)
 
                 num = _to_float(raw)
@@ -200,6 +229,9 @@ def _profile_csv_stream(
 
         nn = non_missing[col]
         top = value_counts[col].most_common(5)
+        if value_overflow[col]['count'] > 0:
+            top.append(('__OTHER__', value_overflow[col]['count']))
+            top = sorted(top, key=lambda x: x[1], reverse=True)[:5]
         top_values = [
             {'value': v, 'count': cnt, 'ratio': round(cnt / row_count, 6) if row_count else 0.0}
             for v, cnt in top
@@ -215,19 +247,21 @@ def _profile_csv_stream(
                 'outlier_ratio': _outlier_ratio(numeric_outlier_samples[col]),
             }
 
-        unique_ratio = round(len(unique_sets[col]) / nn, 6) if nn else 0.0
+        unique_count = _estimate_unique_count(unique_bitmaps[col])
+        unique_ratio = round(min(unique_count, nn) / nn, 6) if nn else 0.0
         dominant_value_ratio = top_values[0]['ratio'] if top_values else 0.0
         profiles[col] = {
             'missing_count': missing[col],
             'missing_ratio': round(missing[col] / row_count, 6) if row_count else 0.0,
             'non_missing_count': nn,
-            'unique_count': len(unique_sets[col]),
+            'unique_count': unique_count,
             'unique_ratio': unique_ratio,
             'dominant_value_ratio': dominant_value_ratio,
             'top_values': top_values,
             'numeric_distribution': numeric_distribution,
             'dtype': dtypes[col],
             'semantic_type': _infer_semantic_type(col, dtypes[col], value_samples[col], unique_ratio),
+            'top_values_capped': value_overflow[col]['count'] > 0,
         }
 
     summary = {
diff --git a/tests/test_analysis.py b/tests/test_analysis.py
index a3a4517..cc88aa7 100644
--- a/tests/test_analysis.py
+++ b/tests/test_analysis.py
@@ -126,3 +126,17 @@ def test_multi_csv_cache_created(tmp_path, monkeypatch):
     result = multi.analyze_multiple_csv([p], "캐시")
     assert result["file_count"] == 1
     assert any((tmp_path / ".cache").glob("*.json"))
+
+
+def test_multi_csv_top_values_capped_marker(monkeypatch, tmp_path):
+    import bitnet_tools.multi_csv as multi
+
+    monkeypatch.setattr(multi, "TOP_VALUE_TRACK_CAP", 3)
+    p = tmp_path / "cardinality.csv"
+    p.write_text("col\na\nb\nc\nd\na\n", encoding="utf-8")
+
+    result = multi.analyze_multiple_csv([p], "카디널리티")
+    prof = result["files"][0]["column_profiles"]["col"]
+
+    assert prof["top_values_capped"] is True
+    assert any(x["value"] == "__OTHER__" for x in prof["top_values"])