Upload batch: migrate_add_language.py, search_index/build_manifest.json, update_indices.py, README.md, search_index/build_stdout.log...

README.md

@@ -0,0 +1,412 @@
+---
+language:
+- en
+license: apache-2.0
+size_categories:
+- 100K<n<1M
+task_categories:
+- audio-classification
+- text-to-audio
+- feature-extraction
+tags:
+- music
+- ai-generated-music
+- audio
+- embeddings
+- search
+- faiss
+- bm25
+- nsfw-detection
+- transcription
+- captioning
+pretty_name: LAION-Tunes
+dataset_info:
+  features:
+  - name: row_id
+    dtype: int64
+  - name: audio_url
+    dtype: string
+  - name: filename
+    dtype: string
+  - name: tar_url
+    dtype: string
+  - name: subset
+    dtype: string
+  - name: title
+    dtype: string
+  - name: tags_text
+    dtype: string
+  - name: mood_text
+    dtype: string
+  - name: has_lyrics
+    dtype: bool
+  - name: genre_tags
+    dtype: string
+  - name: scene_tags
+    dtype: string
+  - name: emotion_tags
+    dtype: string
+  - name: score_coherence
+    dtype: float64
+  - name: score_musicality
+    dtype: float64
+  - name: score_memorability
+    dtype: float64
+  - name: score_clarity
+    dtype: float64
+  - name: score_naturalness
+    dtype: float64
+  - name: score_average
+    dtype: float64
+  - name: play_count
+    dtype: int64
+  - name: upvote_count
+    dtype: int64
+  - name: duration_seconds
+    dtype: float64
+  - name: music_whisper_caption
+    dtype: string
+  - name: parakeet_transcription
+    dtype: string
+  - name: has_caption
+    dtype: bool
+  - name: has_transcription
+    dtype: bool
+  - name: language
+    dtype: string
+  - name: is_instrumental
+    dtype: bool
+  - name: nsfw_gore_sim
+    dtype: float64
+  - name: nsfw_sexual_sim
+    dtype: float64
+  - name: nsfw_hate_sim
+    dtype: float64
+  - name: nsfw_gore_label
+    dtype: string
+  - name: nsfw_sexual_label
+    dtype: string
+  - name: nsfw_hate_label
+    dtype: string
+  - name: nsfw_overall_label
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 908174
+---
+
+# LAION-Tunes
+
+**908,174 AI-generated music tracks** from 5 platforms, annotated with captions, transcriptions, embeddings, aesthetics scores, and NSFW safety labels. Includes a full-text and vector search engine with a web UI.
+
+## Quick Stats
+
+| Metric | Value |
+|--------|-------|
+| Total tracks | 908,174 |
+| Subsets | Mureka (383K), Suno (308K), Udio (115K), Riffusion (99K), Sonauto (3K) |
+| Has caption (Music-Whisper) | 832,944 (91.7%) |
+| Has transcription (Parakeet ASR) | 514,203 (56.6%) |
+| Instrumental | 394,246 (43.4%) |
+| NSFW flagged (very likely + likely) | 12,860 (1.4%) |
+
+## Dataset Description
+
+LAION-Tunes is a curated metadata and annotation dataset derived from [ai-music/ai-music-deduplicated](https://huggingface.co/datasets/ai-music/ai-music-deduplicated), a collection of publicly available AI-generated music from Suno, Udio, Mureka, Riffusion, and Sonauto.
+
+**This dataset does NOT contain audio files.** It contains metadata, annotations, embeddings, and search indices. Audio URLs pointing to the original hosting platforms are included for reference.
+
+### What's Included
+
+For each track:
+- **Metadata**: title, tags, genre, mood, duration, play count, upvote count
+- **Music-Whisper Caption**: AI-generated music description using [laion/music-whisper](https://huggingface.co/laion/music-whisper) (fine-tuned OpenAI Whisper Small)
+- **Parakeet ASR Transcription**: Speech-to-text using [nvidia/parakeet-tdt-0.6b-v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) with word-level timestamps
+- **Sentence Embeddings**: 768-dim embeddings via [google/embeddinggemma-300m](https://huggingface.co/google/embeddinggemma-300m) for tags, captions, transcriptions, lyrics, and moods
+- **Whisper Audio Embeddings**: 768-dim mean-pooled encoder embeddings from Music-Whisper for audio similarity search
+- **Aesthetics Scores**: coherence, musicality, memorability, clarity, naturalness (computed from Music-Whisper)
+- **NSFW Safety Labels**: three-tier classification (very_likely_nsfw / likely_nsfw / likely_sfw) across gore, sexual, and hate speech dimensions
+- **Pre-built Search Indices**: FAISS vector indices and BM25 text indices ready to serve
+
+### Annotation Pipeline
+
+The annotation pipeline processes the original TAR files from ai-music-deduplicated:
+
+1. **Music-Whisper** (`laion/music-whisper`): Generates music captions describing instruments, genre, mood, tempo, etc.
+2. **Parakeet TDT 0.6B** (`nvidia/parakeet-tdt-0.6b-v3`): ASR transcription with word-level timestamps for vocal content
+3. **EmbeddingGemma 300M** (`google/embeddinggemma-300m`): Computes 768-dim sentence embeddings for captions, transcriptions, tags, lyrics, moods
+4. **Whisper Encoder Embeddings**: Mean-pooled encoder hidden states from Music-Whisper for audio fingerprinting/similarity
+5. **NSFW Classification**: Cosine similarity of transcription embeddings against reference prompts for gore/violence, sexual content, and hate speech
+
+## Repository Structure
+
+```
+laion-tunes/
+├── README.md                          # This file
+├── server.py                          # FastAPI search server (main application)
+├── index.html                         # Web UI (dark-mode, single-page app)
+├── build_search_index.py              # Index builder script
+├── update_indices.py                  # Incremental index updater
+├── migrate_add_language.py            # Language detection migration
+├── nsfw_safety_report.html            # Interactive NSFW analysis report
+├── nsfw_analysis_data.json            # Raw NSFW analysis data
+│
+├── public/                            # Annotated metadata parquets (8.3 GB)
+│   ├── mureka_000000.tar.parquet      # One parquet per source TAR file
+│   ├── mureka_000001.tar.parquet
+│   ├── ...
+│   └── udio_000015.tar.parquet        # 159 parquet files total
+│
+├── search_index/                      # Pre-built search indices (16 GB)
+│   ├── metadata.db                    # SQLite database (908K tracks, 2.7 GB)
+│   ├── faiss_tag.index                # FAISS IndexFlatIP - tag embeddings (2.6 GB)
+│   ├── faiss_whisper.index            # FAISS IndexFlatIP - audio embeddings (2.6 GB)
+│   ├── faiss_caption.index            # FAISS IndexFlatIP - caption embeddings (2.3 GB)
+│   ├── faiss_transcription.index      # FAISS IndexFlatIP - transcription embeddings (1.5 GB)
+│   ├── faiss_lyric.index              # FAISS IndexFlatIP - lyric embeddings (1.4 GB)
+│   ├── faiss_mood.index               # FAISS IndexFlatIP - mood embeddings (1.1 GB)
+│   ├── idmap_*.npy                    # Row ID mappings for each FAISS index
+│   ├── bm25_tags.pkl                  # BM25 text index for tags (114 MB)
+│   ├── bm25_caption.pkl               # BM25 text index for captions (609 MB)
+│   └── bm25_transcription.pkl         # BM25 text index for transcriptions (392 MB)
+│
+└── whisper_embeddings/                # Raw Whisper encoder embeddings (1.6 GB)
+    ├── mureka_000000.tar.npz          # One NPZ per source TAR file
+    ├── ...
+    └── udio_000015.tar.npz            # 159 NPZ files total
+```
+
+## Data Format
+
+### Parquet Files (`public/`)
+
+Each parquet file corresponds to one TAR file from the source dataset and contains these columns:
+
+| Column | Type | Description |
+|--------|------|-------------|
+| `filename` | str | Filename within the source TAR |
+| `tar_file` | str | Source TAR filename |
+| `audio_url` | str | Original audio URL (mp3/m4a/ogg) |
+| `subset` | str | Source platform (suno/udio/mureka/riffusion/sonauto) |
+| `title` | str | Track title |
+| `tags` | str | Comma-separated genre/style tags |
+| `mood` | str | Mood tags |
+| `lyrics` | str | Lyrics (if available, from source metadata) |
+| `duration_seconds` | float | Track duration |
+| `play_count` | int | Play count on source platform |
+| `upvote_count` | int | Like/upvote count |
+| `music_whisper_caption` | str | Music-Whisper generated caption |
+| `parakeet_transcription` | str | Parakeet ASR transcription (plain text) |
+| `parakeet_transcription_with_timestamps` | str | ASR with word-level timestamps |
+| `tag_embedding` | list[float] | 768-dim EmbeddingGemma embedding of tags |
+| `caption_embedding` | list[float] | 768-dim EmbeddingGemma embedding of caption |
+| `transcription_embedding` | list[float] | 768-dim EmbeddingGemma embedding of transcription |
+| `lyric_embedding` | list[float] | 768-dim EmbeddingGemma embedding of lyrics |
+| `mood_embedding` | list[float] | 768-dim EmbeddingGemma embedding of mood |
+
+### Whisper Embeddings (`whisper_embeddings/`)
+
+NPZ files containing mean-pooled Whisper encoder hidden states:
+- `embeddings`: float32 array of shape `(N, 768)` - L2-normalized
+- `filenames`: string array of filenames matching the parquet entries
+
+### SQLite Database (`search_index/metadata.db`)
+
+The `tracks` table contains all 908,174 tracks with 34 columns including metadata, aesthetics scores, annotation flags, and NSFW safety labels. The `row_id` column is the primary key used by all FAISS indices.
+
+### FAISS Indices (`search_index/faiss_*.index`)
+
+All indices are `IndexFlatIP` (inner product / cosine similarity for L2-normalized vectors) with 768 dimensions. Each index has a corresponding `idmap_*.npy` that maps FAISS internal indices to SQLite `row_id` values.
+
+| Index | Vectors | Description |
+|-------|---------|-------------|
+| `faiss_tag` | 908,241 | Tag text embeddings |
+| `faiss_whisper` | 908,174 | Audio encoder embeddings (music similarity) |
+| `faiss_caption` | 798,858 | Music-Whisper caption embeddings |
+| `faiss_transcription` | 511,610 | ASR transcription embeddings |
+| `faiss_lyric` | 479,313 | Lyrics embeddings |
+| `faiss_mood` | 383,616 | Mood text embeddings |
+
+## NSFW Safety Labels
+
+Each track has NSFW safety scores and labels across three dimensions:
+
+| Dimension | Strict Threshold | Moderate Threshold | Very Likely NSFW | Likely NSFW |
+|-----------|-----------------|-------------------|-----------------|-------------|
+| Gore/Violence | >= 0.3779 | >= 0.3540 | 2,437 (0.27%) | 2,293 (0.25%) |
+| Sexual Content | >= 0.3584 | >= 0.3234 | 3,367 (0.37%) | 2,689 (0.30%) |
+| Hate Speech | >= 0.3633 | >= 0.3382 | 2,786 (0.31%) | 2,316 (0.26%) |
+| **Overall (conservative)** | - | - | **6,762 (0.74%)** | **6,098 (0.67%)** |
+
+- `very_likely_nsfw`: cosine similarity above strict threshold
+- `likely_nsfw`: between strict and moderate thresholds
+- `likely_sfw`: below moderate threshold
+- `nsfw_overall_label`: conservative (worst label across all three dimensions wins)
+
+The raw cosine similarity scores (`nsfw_gore_sim`, `nsfw_sexual_sim`, `nsfw_hate_sim`) are stored so you can apply your own thresholds.
+
+## Running the Search Server
+
+### Prerequisites
+
+```bash
+pip install fastapi uvicorn faiss-cpu numpy pandas sentence-transformers torch scipy tqdm
+```
+
+### Option 1: With HF Text Embeddings Inference (Recommended)
+
+TEI provides fast CPU-based embedding serving (~25ms per query vs ~430ms with Python):
+
+```bash
+# Start TEI (requires Docker)
+docker run -d --name tei-embeddings \
+  -p 8090:80 \
+  -e HF_TOKEN=your_token \
+  ghcr.io/huggingface/text-embeddings-inference:cpu-latest \
+  --model-id google/embeddinggemma-300m \
+  --max-batch-requests 4
+
+# Start the server with TEI backend
+python server.py --port 7860 --gpu 0 --tei-url http://localhost:8090
+```
+
+### Option 2: Direct Python Inference
+
+```bash
+python server.py --port 7860 --gpu 0
+```
+
+This loads EmbeddingGemma 300M and Music-Whisper encoder into Python directly.
+
+### Server Flags
+
+| Flag | Default | Description |
+|------|---------|-------------|
+| `--port` | 7860 | HTTP port |
+| `--host` | 0.0.0.0 | Bind address |
+| `--gpu` | 0 | GPU ID for Whisper encoder |
+| `--tei-url` | None | TEI server URL for text embeddings |
+| `--no-whisper` | False | Skip loading Whisper encoder (disables audio similarity search) |
+
+### What Loads at Startup
+
+1. **6 FAISS indices** (tag, whisper, caption, transcription, lyric, mood)
+2. **3 BM25 indices** (tags, caption, transcription)
+3. **SQLite database** (908K tracks)
+4. **Text embedder**: TEI backend or Python SentenceTransformer
+5. **Music-Whisper encoder** (on GPU): for audio upload similarity search
+
+Total memory: ~20 GB RAM + ~200 MB GPU VRAM
+
+## Search API
+
+### `POST /api/search`
+
+Main search endpoint combining vector search, BM25, and metadata filtering.
+
+```json
+{
+  "query": "upbeat electronic dance music",
+  "search_fields": ["tag", "caption"],
+  "bm25_fields": ["tags", "caption"],
+  "top_k": 50,
+  "bm25_weight": 0.3,
+  "min_score": 4.0,
+  "max_score": 10.0,
+  "subsets": ["suno", "udio"],
+  "nsfw_filter": "sfw_only",
+  "stage2_enabled": true,
+  "stage2_field": "caption",
+  "stage2_top_k": 200
+}
+```
+
+**Search fields** (FAISS vector search):
+- `tag`: search by music tags/genres
+- `caption`: search by Music-Whisper captions
+- `transcription`: search by ASR transcriptions
+- `lyric`: search by lyrics content
+- `mood`: search by mood descriptors
+- `whisper`: search by audio embedding similarity
+
+**BM25 fields** (text search):
+- `tags`, `caption`, `transcription`
+
+**Filters**:
+- `min_score` / `max_score`: aesthetics score range (0-10)
+- `subsets`: list of source platforms
+- `nsfw_filter`: `"sfw_only"` or `"nsfw_only"` or `null` for all
+
+**Two-stage search**: First retrieves `stage2_top_k` candidates, then re-ranks by a second field.
+
+### `POST /api/search_by_audio`
+
+Upload an audio file to find similar tracks by audio fingerprint.
+
+```bash
+curl -X POST http://localhost:7860/api/search_by_audio \
+  -F "file=@song.mp3" \
+  -F "top_k=20" \
+  -F "nsfw_filter=sfw_only"
+```
+
+### `POST /api/search_similar`
+
+Find tracks similar to an existing track by row_id.
+
+```json
+{
+  "row_id": 12345,
+  "field": "whisper",
+  "top_k": 20
+}
+```
+
+### `GET /api/stats`
+
+Returns dataset statistics (total tracks, index sizes, NSFW counts).
+
+## Building the Index from Scratch
+
+If you want to rebuild the search indices from the parquet files:
+
+```bash
+python build_search_index.py --force
+```
+
+This reads all parquets from `public/` (and optionally `private/` for additional embeddings), builds the SQLite database, FAISS indices, and BM25 indices. Takes ~30 minutes on a modern machine.
+
+## Source Data
+
+The original audio data comes from [ai-music/ai-music-deduplicated](https://huggingface.co/datasets/ai-music/ai-music-deduplicated), organized by platform:
+
+| Platform | Tracks | Audio Format | Notable Fields |
+|----------|--------|-------------|----------------|
+| Mureka | 383,549 | MP3 | genres, moods, model version |
+| Suno | 307,539 | MP3 | tags (in metadata), prompt, model_name, explicit flag |
+| Udio | 115,140 | MP3 | tags (array), lyrics, prompt, likes, plays |
+| Riffusion | 99,228 | M4A | sound (style description), lyrics_timestamped, conditions |
+| Sonauto | 2,718 | OGG | tags (array), description, keyword |
+
+## Models Used
+
+| Model | Purpose | Output |
+|-------|---------|--------|
+| [laion/music-whisper](https://huggingface.co/laion/music-whisper) | Music captioning + audio embeddings | Text caption + 768-dim encoder embedding |
+| [nvidia/parakeet-tdt-0.6b-v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) | ASR transcription | Text + word-level timestamps |
+| [google/embeddinggemma-300m](https://huggingface.co/google/embeddinggemma-300m) | Text sentence embeddings | 768-dim L2-normalized vectors |
+
+## License
+
+Apache 2.0
+
+## Citation
+
+```bibtex
+@misc{laion-tunes-2025,
+  title={LAION-Tunes: Annotated AI Music Search Dataset},
+  author={LAION},
+  year={2025},
+  url={https://huggingface.co/datasets/laion/laion-tunes}
+}
+```

build_search_index.py

@@ -0,0 +1,609 @@
+#!/usr/bin/env python3
+"""
+LAION-Tunes Search Index Builder
+=================================
+Reads annotated parquet files from two sources, merges them, and builds:
+  1. SQLite metadata database (tracks table with all searchable fields)
+  2. FAISS vector indices (tag, lyric, mood, caption, transcription embeddings)
+  3. BM25 text search indices (tags, caption, transcription, hashed lyrics)
+
+Data Sources:
+  - private/ parquets: 908K rows with tag/lyric/mood embeddings, aesthetics scores
+  - output/ parquets: annotation pipeline output with music_whisper_caption, parakeet ASR
+
+Usage:
+    python build_search_index.py [--force]  # --force rebuilds from scratch
+
+Output: search_index/ directory with all index files.
+Resumable: tracks which parquets have been processed in build_manifest.json.
+"""
+
+import os
+import sys
+import json
+import glob
+import time
+import hmac
+import hashlib
+import pickle
+import re
+import sqlite3
+import logging
+from pathlib import Path
+from datetime import timedelta
+from collections import defaultdict
+
+import numpy as np
+import pandas as pd
+import faiss
+from scipy import sparse
+from tqdm import tqdm
+
+# ── Paths ────────────────────────────────────────────────────────────────────
+BASE_DIR = Path(__file__).parent
+PRIVATE_DIR = BASE_DIR / "private"
+OUTPUT_DIR = Path("/home/deployer/laion/music/output")
+INDEX_DIR = BASE_DIR / "search_index"
+MANIFEST_PATH = INDEX_DIR / "build_manifest.json"
+
+# BM25 lyrics hashing secret - change this to your own secret
+LYRICS_HMAC_KEY = b"laion-tunes-search-2026-secret-key"
+
+# ── Logging ──────────────────────────────────────────────────────────────────
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s [%(levelname)s] %(message)s",
+    datefmt="%H:%M:%S",
+    handlers=[
+        logging.StreamHandler(sys.stdout),
+        logging.FileHandler(str(INDEX_DIR / "build.log") if INDEX_DIR.exists() else "/dev/null", mode="a"),
+    ],
+)
+log = logging.getLogger("index_builder")
+
+
+# ── BM25 Implementation ─────────────────────────────────────────────────────
+class BM25Index:
+    """Lightweight BM25-Okapi index using sparse matrices for fast scoring."""
+
+    def __init__(self, k1=1.5, b=0.75):
+        self.k1 = k1
+        self.b = b
+        self.vocab = {}          # token -> vocab_id
+        self.idf = None          # numpy array, shape (vocab_size,)
+        self.tf_matrix = None    # scipy CSR, shape (n_docs, vocab_size)
+        self.doc_lens = None     # numpy array, shape (n_docs,)
+        self.avg_dl = 0.0
+        self.row_ids = []        # maps doc index -> SQLite row_id
+
+    def build(self, documents, row_ids):
+        """Build index from list of (tokenized_doc, row_id) pairs.
+        documents: list of list of strings (already tokenized)
+        row_ids: list of ints (SQLite row_ids)
+        """
+        self.row_ids = list(row_ids)
+        n_docs = len(documents)
+        if n_docs == 0:
+            return
+
+        # Build vocabulary
+        doc_freqs = defaultdict(int)  # token -> number of docs containing it
+        for doc in documents:
+            seen = set()
+            for token in doc:
+                if token not in self.vocab:
+                    self.vocab[token] = len(self.vocab)
+                if token not in seen:
+                    doc_freqs[token] += 1
+                    seen.add(token)
+
+        vocab_size = len(self.vocab)
+
+        # Build sparse TF matrix
+        rows, cols, data = [], [], []
+        doc_lens = np.zeros(n_docs, dtype=np.float32)
+        for doc_idx, doc in enumerate(documents):
+            tf = defaultdict(int)
+            for token in doc:
+                tf[self.vocab[token]] += 1
+            doc_lens[doc_idx] = len(doc)
+            for vid, count in tf.items():
+                rows.append(doc_idx)
+                cols.append(vid)
+                data.append(count)
+
+        self.tf_matrix = sparse.csr_matrix(
+            (np.array(data, dtype=np.float32), (rows, cols)),
+            shape=(n_docs, vocab_size),
+        )
+        self.doc_lens = doc_lens
+        self.avg_dl = float(doc_lens.mean()) if n_docs > 0 else 1.0
+
+        # Compute IDF: log((N - df + 0.5) / (df + 0.5) + 1)
+        self.idf = np.zeros(vocab_size, dtype=np.float32)
+        for token, vid in self.vocab.items():
+            df = doc_freqs[token]
+            self.idf[vid] = np.log((n_docs - df + 0.5) / (df + 0.5) + 1.0)
+
+    def search(self, query_tokens, top_k=50):
+        """Score all documents against query tokens. Returns (row_ids, scores)."""
+        if self.tf_matrix is None or len(query_tokens) == 0:
+            return np.array([], dtype=np.int64), np.array([], dtype=np.float32)
+
+        # Map query tokens to vocab ids
+        qvids = [self.vocab[t] for t in query_tokens if t in self.vocab]
+        if not qvids:
+            return np.array([], dtype=np.int64), np.array([], dtype=np.float32)
+
+        n_docs = self.tf_matrix.shape[0]
+        scores = np.zeros(n_docs, dtype=np.float32)
+
+        for vid in qvids:
+            tf_col = self.tf_matrix[:, vid].toarray().ravel()
+            idf_val = self.idf[vid]
+            # BM25 formula
+            numerator = tf_col * (self.k1 + 1)
+            denominator = tf_col + self.k1 * (1 - self.b + self.b * self.doc_lens / self.avg_dl)
+            scores += idf_val * (numerator / (denominator + 1e-8))
+
+        # Top-k
+        k = min(top_k, n_docs)
+        top_idx = np.argpartition(scores, -k)[-k:]
+        top_idx = top_idx[np.argsort(scores[top_idx])[::-1]]
+        top_idx = top_idx[scores[top_idx] > 0]
+
+        result_row_ids = np.array([self.row_ids[i] for i in top_idx], dtype=np.int64)
+        result_scores = scores[top_idx]
+        return result_row_ids, result_scores
+
+
+# ── Tokenizer ────────────────────────────────────────────────────────────────
+_TOKEN_RE = re.compile(r"[a-z0-9]+")
+
+def tokenize(text):
+    """Lowercase, split on non-alphanumeric, keep tokens >= 2 chars."""
+    if not text or not isinstance(text, str):
+        return []
+    return [t for t in _TOKEN_RE.findall(text.lower()) if len(t) >= 2]
+
+
+def tokenize_and_hash(text, secret_key=LYRICS_HMAC_KEY):
+    """Tokenize then HMAC-SHA256 hash each token (privacy-preserving)."""
+    tokens = tokenize(text)
+    return [hmac.new(secret_key, t.encode(), hashlib.sha256).hexdigest()[:16] for t in tokens]
+
+
+# ── Title / Lyrics extraction from metadata_json ─────────────────────────────
+def extract_title(metadata_json_str):
+    """Extract title from the raw metadata JSON string."""
+    try:
+        meta = json.loads(metadata_json_str)
+        return meta.get("title", "") or ""
+    except Exception:
+        return ""
+
+
+def extract_lyrics(metadata_json_str):
+    """Extract lyrics text from metadata JSON. Returns empty string if not available.
+    Suno: metadata.prompt (contains lyrics/prompt)
+    Udio: lyrics field
+    Riffusion: lyrics field
+    Others: not available
+    """
+    try:
+        meta = json.loads(metadata_json_str)
+        # Suno: lyrics are in metadata.prompt
+        nested = meta.get("metadata", {})
+        if isinstance(nested, dict) and nested.get("prompt"):
+            return str(nested["prompt"])
+        # Udio / Riffusion / Sonauto
+        if meta.get("lyrics"):
+            return str(meta["lyrics"])
+        return ""
+    except Exception:
+        return ""
+
+
+def extract_duration(metadata_json_str):
+    """Extract duration from metadata JSON."""
+    try:
+        meta = json.loads(metadata_json_str)
+        # Suno: metadata.duration
+        nested = meta.get("metadata", {})
+        if isinstance(nested, dict) and nested.get("duration"):
+            return float(nested["duration"])
+        # Udio: duration
+        if meta.get("duration"):
+            return float(meta["duration"])
+        # Riffusion: duration_s
+        if meta.get("duration_s"):
+            return float(meta["duration_s"])
+        # Mureka: duration in milliseconds
+        if meta.get("duration in milliseconds"):
+            return float(meta["duration in milliseconds"]) / 1000.0
+        return None
+    except Exception:
+        return None
+
+
+def extract_subset(tar_url):
+    """Extract subset name from tar_url like '.../suno/suno_000000.tar'."""
+    if not tar_url:
+        return "unknown"
+    parts = tar_url.split("/")
+    for p in parts:
+        if p in ("suno", "udio", "mureka", "riffusion", "sonauto"):
+            return p
+    return "unknown"
+
+
+# ── Index building ───────────────────────────────────────────────────────────
+def load_manifest():
+    """Load or create build manifest for resumability."""
+    if MANIFEST_PATH.exists():
+        with open(MANIFEST_PATH) as f:
+            return json.load(f)
+    return {"processed_private": [], "processed_output": [], "total_rows": 0}
+
+
+def save_manifest(manifest):
+    with open(MANIFEST_PATH, "w") as f:
+        json.dump(manifest, f, indent=2)
+
+
+def get_output_filename(private_filename):
+    """Map private parquet filename to expected output parquet filename.
+    'suno_000000.tar.parquet' -> 'suno_suno_000000.parquet'
+    """
+    base = private_filename.replace(".tar.parquet", "")
+    parts = base.split("_", 1)
+    if len(parts) == 2:
+        subset, num = parts
+        return f"{subset}_{subset}_{num}.parquet"
+    return None
+
+
+def init_database(db_path):
+    """Create SQLite database with tracks table."""
+    conn = sqlite3.connect(str(db_path))
+    conn.execute("PRAGMA journal_mode=WAL")
+    conn.execute("PRAGMA synchronous=NORMAL")
+    conn.execute("""
+        CREATE TABLE IF NOT EXISTS tracks (
+            row_id        INTEGER PRIMARY KEY,
+            audio_url     TEXT NOT NULL,
+            filename      TEXT,
+            tar_url       TEXT,
+            subset        TEXT NOT NULL,
+            title         TEXT,
+            tags_text     TEXT,
+            mood_text     TEXT,
+            has_lyrics    INTEGER NOT NULL DEFAULT 0,
+            genre_tags    TEXT,
+            scene_tags    TEXT,
+            emotion_tags  TEXT,
+            score_coherence    REAL,
+            score_musicality   REAL,
+            score_memorability REAL,
+            score_clarity      REAL,
+            score_naturalness  REAL,
+            score_average      REAL,
+            play_count    INTEGER DEFAULT 0,
+            upvote_count  INTEGER DEFAULT 0,
+            duration_seconds   REAL,
+            music_whisper_caption  TEXT,
+            parakeet_transcription TEXT,
+            has_caption   INTEGER NOT NULL DEFAULT 0,
+            has_transcription INTEGER NOT NULL DEFAULT 0
+        )
+    """)
+    conn.execute("CREATE INDEX IF NOT EXISTS idx_score_avg ON tracks(score_average)")
+    conn.execute("CREATE INDEX IF NOT EXISTS idx_play ON tracks(play_count)")
+    conn.execute("CREATE INDEX IF NOT EXISTS idx_upvote ON tracks(upvote_count)")
+    conn.execute("CREATE INDEX IF NOT EXISTS idx_subset ON tracks(subset)")
+    conn.execute("CREATE UNIQUE INDEX IF NOT EXISTS idx_audio_url ON tracks(audio_url)")
+    conn.commit()
+    return conn
+
+
+def build_all(force=False):
+    """Main index building function."""
+    INDEX_DIR.mkdir(parents=True, exist_ok=True)
+
+    # Re-init logging file handler now that dir exists
+    fh = logging.FileHandler(str(INDEX_DIR / "build.log"), mode="a")
+    fh.setFormatter(logging.Formatter("%(asctime)s [%(levelname)s] %(message)s", datefmt="%H:%M:%S"))
+    log.addHandler(fh)
+
+    log.info("=" * 60)
+    log.info("LAION-Tunes Search Index Builder")
+    log.info("=" * 60)
+
+    manifest = load_manifest() if not force else {"processed_private": [], "processed_output": [], "total_rows": 0}
+    if force:
+        log.info("Force rebuild requested - starting fresh")
+
+    # List all private parquets
+    private_files = sorted(PRIVATE_DIR.glob("*.parquet"))
+    log.info(f"Found {len(private_files)} private parquets")
+
+    # List all output parquets (annotation pipeline results)
+    output_files = {Path(f).name: f for f in glob.glob(str(OUTPUT_DIR / "*.parquet"))}
+    log.info(f"Found {len(output_files)} output (annotation) parquets")
+
+    # Init or open SQLite
+    db_path = INDEX_DIR / "metadata.db"
+    if force and db_path.exists():
+        db_path.unlink()
+    conn = init_database(db_path)
+
+    # Get current row count
+    row_id_offset = conn.execute("SELECT COALESCE(MAX(row_id), -1) + 1 FROM tracks").fetchone()[0]
+    log.info(f"Starting row_id offset: {row_id_offset}")
+
+    # FAISS indices - load existing or create new
+    faiss_indices = {}
+    id_maps = {}
+    embedding_fields = ["tag", "lyric", "mood", "caption", "transcription"]
+
+    for field in embedding_fields:
+        idx_path = INDEX_DIR / f"faiss_{field}.index"
+        map_path = INDEX_DIR / f"idmap_{field}.npy"
+        if not force and idx_path.exists() and map_path.exists():
+            faiss_indices[field] = faiss.read_index(str(idx_path))
+            id_maps[field] = list(np.load(str(map_path)))
+            log.info(f"  Loaded existing FAISS {field}: {faiss_indices[field].ntotal} vectors")
+        else:
+            faiss_indices[field] = faiss.IndexFlatIP(768)
+            id_maps[field] = []
+            log.info(f"  Created new FAISS {field} index")
+
+    # BM25 data collectors - always rebuild from scratch (fast enough)
+    bm25_data = {
+        "tags": {"docs": [], "row_ids": []},
+        "caption": {"docs": [], "row_ids": []},
+        "transcription": {"docs": [], "row_ids": []},
+        "lyrics_hashed": {"docs": [], "row_ids": []},
+    }
+
+    # If resuming, we need to rebuild BM25 from the entire DB
+    if row_id_offset > 0 and not force:
+        log.info("Rebuilding BM25 data from existing database...")
+        cursor = conn.execute(
+            "SELECT row_id, tags_text, music_whisper_caption, parakeet_transcription FROM tracks"
+        )
+        for row in cursor:
+            rid, tags, caption, transcript = row
+            if tags:
+                bm25_data["tags"]["docs"].append(tokenize(tags))
+                bm25_data["tags"]["row_ids"].append(rid)
+            if caption:
+                bm25_data["caption"]["docs"].append(tokenize(caption))
+                bm25_data["caption"]["row_ids"].append(rid)
+            if transcript:
+                bm25_data["transcription"]["docs"].append(tokenize(transcript))
+                bm25_data["transcription"]["row_ids"].append(rid)
+        log.info(f"  BM25 tags: {len(bm25_data['tags']['docs'])}, "
+                 f"caption: {len(bm25_data['caption']['docs'])}, "
+                 f"transcription: {len(bm25_data['transcription']['docs'])}")
+
+    # Process private parquets
+    files_to_process = [f for f in private_files if f.name not in manifest["processed_private"]]
+    log.info(f"Processing {len(files_to_process)} new private parquets "
+             f"(skipping {len(private_files) - len(files_to_process)} already done)")
+
+    t0 = time.time()
+    total_new = 0
+
+    for file_idx, pf in enumerate(tqdm(files_to_process, desc="Building index")):
+        df = pd.read_parquet(pf)
+        n = len(df)
+
+        # Try to find matching output parquet for annotation data
+        out_name = get_output_filename(pf.name)
+        out_df = None
+        if out_name and out_name in output_files:
+            try:
+                out_df = pd.read_parquet(output_files[out_name])
+                out_df = out_df.set_index("audio_url")
+            except Exception as e:
+                log.warning(f"  Failed to load output {out_name}: {e}")
+
+        # Prepare batch data for SQLite insert
+        insert_rows = []
+        batch_embeddings = {field: ([], []) for field in embedding_fields}  # (vectors, row_ids)
+
+        for i in range(n):
+            row = df.iloc[i]
+            row_id = row_id_offset + total_new
+            audio_url = str(row.get("audio_url", ""))
+            if not audio_url:
+                continue
+
+            metadata_json = str(row.get("metadata_json", "{}"))
+            subset = extract_subset(str(row.get("tar_url", "")))
+            title = extract_title(metadata_json)
+            tags_text = str(row.get("tags_text", "")) or ""
+            mood_text = str(row.get("mood_text", "")) if pd.notna(row.get("mood_text")) else ""
+            has_lyrics = bool(row.get("has_lyrics", False))
+            duration = extract_duration(metadata_json)
+
+            # List columns -> JSON strings
+            genre_tags = row.get("genre_tags", [])
+            scene_tags = row.get("scene_tags", [])
+            emotion_tags = row.get("emotion_tags", [])
+            def to_json_list(val):
+                if isinstance(val, np.ndarray):
+                    return json.dumps(val.tolist())
+                if isinstance(val, list):
+                    return json.dumps(val)
+                return "[]"
+            genre_str = to_json_list(genre_tags)
+            scene_str = to_json_list(scene_tags)
+            emotion_str = to_json_list(emotion_tags)
+
+            # Scores
+            score_avg = float(row.get("score_average", 0)) if pd.notna(row.get("score_average")) else None
+            score_coh = float(row.get("score_Coherence", 0)) if pd.notna(row.get("score_Coherence")) else None
+            score_mus = float(row.get("score_Musicality", 0)) if pd.notna(row.get("score_Musicality")) else None
+            score_mem = float(row.get("score_Memorability", 0)) if pd.notna(row.get("score_Memorability")) else None
+            score_cla = float(row.get("score_Clarity", 0)) if pd.notna(row.get("score_Clarity")) else None
+            score_nat = float(row.get("score_Naturalness", 0)) if pd.notna(row.get("score_Naturalness")) else None
+            play_count = int(row.get("play_count", 0) or 0)
+            upvote_count = int(row.get("upvote_count", 0) or 0)
+
+            # Merge annotation data if available
+            caption = ""
+            transcription = ""
+            out_duration = duration
+            if out_df is not None and audio_url in out_df.index:
+                out_row = out_df.loc[audio_url]
+                if isinstance(out_row, pd.DataFrame):
+                    out_row = out_row.iloc[0]
+                caption = str(out_row.get("music_whisper_caption", "")) if pd.notna(out_row.get("music_whisper_caption")) else ""
+                transcription = str(out_row.get("parakeet_transcription", "")) if pd.notna(out_row.get("parakeet_transcription")) else ""
+                if not title:
+                    title = str(out_row.get("title", "")) if pd.notna(out_row.get("title")) else ""
+                if out_duration is None and pd.notna(out_row.get("duration_seconds")):
+                    out_duration = float(out_row["duration_seconds"])
+
+            insert_rows.append((
+                row_id, audio_url, str(row.get("filename", "")), str(row.get("tar_url", "")),
+                subset, title, tags_text, mood_text, int(has_lyrics),
+                genre_str, scene_str, emotion_str,
+                score_coh, score_mus, score_mem, score_cla, score_nat, score_avg,
+                play_count, upvote_count, out_duration,
+                caption, transcription,
+                1 if caption else 0, 1 if transcription else 0,
+            ))
+
+            # Collect embeddings
+            for field, col_name in [
+                ("tag", "tag_embedding"), ("lyric", "lyric_embedding"),
+                ("mood", "mood_embedding"),
+            ]:
+                emb = row.get(col_name)
+                if emb is not None and isinstance(emb, np.ndarray) and len(emb) == 768:
+                    vec = emb.astype(np.float32)
+                    norm = np.linalg.norm(vec)
+                    if norm > 0:
+                        vec /= norm
+                    batch_embeddings[field][0].append(vec)
+                    batch_embeddings[field][1].append(row_id)
+
+            # Caption/transcription embeddings from output parquet
+            if out_df is not None and audio_url in out_df.index:
+                out_row_data = out_df.loc[audio_url]
+                if isinstance(out_row_data, pd.DataFrame):
+                    out_row_data = out_row_data.iloc[0]
+                for field, col_name in [("caption", "caption_embedding"), ("transcription", "transcription_embedding")]:
+                    emb = out_row_data.get(col_name)
+                    if emb is not None and isinstance(emb, np.ndarray) and len(emb) == 768:
+                        vec = emb.astype(np.float32)
+                        norm = np.linalg.norm(vec)
+                        if norm > 0:
+                            vec /= norm
+                        batch_embeddings[field][0].append(vec)
+                        batch_embeddings[field][1].append(row_id)
+
+            # BM25 data
+            if tags_text:
+                bm25_data["tags"]["docs"].append(tokenize(tags_text))
+                bm25_data["tags"]["row_ids"].append(row_id)
+            if caption:
+                bm25_data["caption"]["docs"].append(tokenize(caption))
+                bm25_data["caption"]["row_ids"].append(row_id)
+            if transcription:
+                bm25_data["transcription"]["docs"].append(tokenize(transcription))
+                bm25_data["transcription"]["row_ids"].append(row_id)
+
+            # Hashed lyrics BM25
+            lyrics_text = extract_lyrics(metadata_json)
+            if lyrics_text and len(lyrics_text) > 10:
+                hashed_tokens = tokenize_and_hash(lyrics_text)
+                if hashed_tokens:
+                    bm25_data["lyrics_hashed"]["docs"].append(hashed_tokens)
+                    bm25_data["lyrics_hashed"]["row_ids"].append(row_id)
+
+            total_new += 1
+
+        # Batch insert into SQLite
+        conn.executemany(
+            "INSERT OR IGNORE INTO tracks VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)",
+            insert_rows,
+        )
+        conn.commit()
+
+        # Batch add to FAISS indices
+        for field in embedding_fields:
+            vecs, rids = batch_embeddings[field]
+            if vecs:
+                arr = np.stack(vecs)
+                faiss_indices[field].add(arr)
+                id_maps[field].extend(rids)
+
+        manifest["processed_private"].append(pf.name)
+        manifest["total_rows"] = row_id_offset + total_new
+        save_manifest(manifest)
+
+        elapsed = time.time() - t0
+        rate = total_new / elapsed if elapsed > 0 else 0
+        remaining_files = len(files_to_process) - file_idx - 1
+        if rate > 0:
+            est_remaining = remaining_files * (n / rate)
+            eta = timedelta(seconds=int(est_remaining))
+        else:
+            eta = "?"
+
+        if (file_idx + 1) % 10 == 0 or (file_idx + 1) == len(files_to_process):
+            log.info(f"  [{file_idx+1}/{len(files_to_process)}] {pf.name}: "
+                     f"{n} rows, total={row_id_offset + total_new:,}, "
+                     f"{rate:.0f} rows/s, ETA: {eta}")
+
+    log.info(f"\nInserted {total_new:,} new rows (total: {row_id_offset + total_new:,})")
+
+    # Save FAISS indices
+    log.info("Saving FAISS indices...")
+    for field in embedding_fields:
+        n_vecs = faiss_indices[field].ntotal
+        if n_vecs > 0:
+            faiss.write_index(faiss_indices[field], str(INDEX_DIR / f"faiss_{field}.index"))
+            np.save(str(INDEX_DIR / f"idmap_{field}.npy"), np.array(id_maps[field], dtype=np.int64))
+            log.info(f"  {field}: {n_vecs:,} vectors")
+        else:
+            log.info(f"  {field}: 0 vectors (skipped)")
+
+    # Build and save BM25 indices
+    log.info("Building BM25 indices...")
+    for bm25_name, bdata in bm25_data.items():
+        if bdata["docs"]:
+            idx = BM25Index()
+            idx.build(bdata["docs"], bdata["row_ids"])
+            with open(INDEX_DIR / f"bm25_{bm25_name}.pkl", "wb") as f:
+                pickle.dump(idx, f, protocol=pickle.HIGHEST_PROTOCOL)
+            log.info(f"  {bm25_name}: {len(bdata['docs']):,} documents, "
+                     f"{len(idx.vocab):,} unique tokens")
+        else:
+            log.info(f"  {bm25_name}: 0 documents (skipped)")
+
+    # Final stats
+    conn.close()
+    save_manifest(manifest)
+    elapsed = time.time() - t0
+    log.info(f"\nDone in {timedelta(seconds=int(elapsed))}")
+    log.info(f"Index directory: {INDEX_DIR}")
+    log.info(f"Total tracks: {row_id_offset + total_new:,}")
+
+    # Print summary
+    for field in embedding_fields:
+        n = faiss_indices[field].ntotal
+        size_mb = n * 768 * 4 / 1024 / 1024
+        log.info(f"  FAISS {field}: {n:,} vectors ({size_mb:.0f} MB)")
+
+    db_size = os.path.getsize(db_path) / 1024 / 1024
+    log.info(f"  SQLite: {db_size:.0f} MB")
+
+
+if __name__ == "__main__":
+    force = "--force" in sys.argv
+    build_all(force=force)

index.html

@@ -0,0 +1,1184 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>LAION-Tunes Search</title>
+<style>
+/* ── Reset & Base ─────────────────────────────────────────────── */
+*, *::before, *::after { box-sizing: border-box; margin: 0; padding: 0; }
+body {
+    font-family: 'Segoe UI', -apple-system, BlinkMacSystemFont, sans-serif;
+    background: #0f1419;
+    color: #e7e9ea;
+    line-height: 1.5;
+    min-height: 100vh;
+}
+
+/* ── Header ──────────────────────────────────────────────────── */
+.header {
+    position: sticky; top: 0; z-index: 100;
+    background: linear-gradient(135deg, #1a1f2e 0%, #16213e 100%);
+    border-bottom: 2px solid #2d4a7a;
+    padding: 16px 24px 12px;
+    box-shadow: 0 4px 20px rgba(0,0,0,0.5);
+}
+.header h1 {
+    font-size: 22px; font-weight: 700; color: #fff; margin-bottom: 12px;
+}
+.header h1 span { color: #64b5f6; }
+
+/* ── Search Bar ──────────────────────────────────────────────── */
+.search-row { display: flex; gap: 10px; margin-bottom: 10px; }
+.search-input {
+    flex: 1; background: #1e2a3a; color: #e0e0e0;
+    border: 1px solid #37474f; border-radius: 8px;
+    padding: 10px 16px; font-size: 15px; outline: none;
+}
+.search-input:focus { border-color: #4fc3f7; }
+.search-input::placeholder { color: #607d8b; }
+.btn-search {
+    background: #2196F3; color: white; border: none;
+    border-radius: 8px; padding: 10px 24px;
+    font-size: 15px; font-weight: 600; cursor: pointer;
+    transition: background 0.2s;
+}
+.btn-search:hover { background: #1976D2; }
+
+/* ── Controls ────────────────────────────────────────────────── */
+.controls { display: flex; flex-wrap: wrap; gap: 10px; align-items: center; margin-bottom: 8px; }
+.control-group { display: flex; align-items: center; gap: 5px; }
+.control-group label { font-size: 12px; color: #90a4ae; font-weight: 600; white-space: nowrap; }
+.control-group select, .control-group input[type="number"] {
+    background: #1e2a3a; color: #e0e0e0;
+    border: 1px solid #37474f; border-radius: 6px;
+    padding: 5px 8px; font-size: 12px; outline: none;
+}
+.control-group select:focus, .control-group input:focus { border-color: #4fc3f7; }
+.control-group input[type="range"] {
+    background: transparent; width: 80px; accent-color: #4fc3f7;
+}
+
+/* ── Filter Chips ────────────────────────────────────────────── */
+.filters { display: flex; flex-wrap: wrap; gap: 6px; align-items: center; margin-bottom: 6px; }
+.filter-section { display: flex; flex-wrap: wrap; gap: 4px; align-items: center; margin-right: 12px; }
+.filter-section-label { font-size: 11px; color: #607d8b; font-weight: 600; margin-right: 2px; }
+.chip {
+    display: inline-flex; align-items: center; gap: 3px;
+    background: #1e2a3a; border: 1px solid #37474f;
+    border-radius: 14px; padding: 2px 10px; font-size: 11px; cursor: pointer;
+    transition: all 0.15s; user-select: none;
+}
+.chip:hover { border-color: #4fc3f7; }
+.chip.active { background: #1a3a5c; border-color: #4fc3f7; color: #4fc3f7; }
+.chip input[type="checkbox"] { display: none; }
+.aes-val { color: #4fc3f7; font-weight: 700; min-width: 24px; text-align: center; font-size: 11px; }
+
+/* ── Language Panel ──────────────────────────────────────────── */
+.lang-panel {
+    display: flex; flex-wrap: wrap; gap: 4px; align-items: center;
+    max-width: 100%; overflow: hidden;
+}
+.lang-chip {
+    display: inline-flex; align-items: center;
+    background: #1e2a3a; border: 1px solid #2a3a4a;
+    border-radius: 12px; padding: 1px 8px; font-size: 10px; cursor: pointer;
+    transition: all 0.15s; user-select: none; color: #90a4ae;
+}
+.lang-chip:hover { border-color: #4fc3f7; }
+.lang-chip.active { background: #1a3a5c; border-color: #4fc3f7; color: #4fc3f7; }
+.lang-chip .cnt { color: #546e7a; margin-left: 3px; }
+
+/* ── Stage 2 Panel ──────────────────────────────────────────── */
+.stage2-panel {
+    margin-top: 8px; padding: 10px 14px;
+    background: rgba(255,152,0,0.06); border: 1px solid #5d4037;
+    border-radius: 8px; display: none;
+}
+.stage2-panel.active { display: block; }
+.stage2-header {
+    display: flex; align-items: center; gap: 8px; margin-bottom: 8px;
+}
+.stage2-header .label { font-size: 12px; font-weight: 700; color: #ffb74d; }
+.stage2-row { display: flex; gap: 10px; margin-bottom: 6px; }
+.stage2-input {
+    flex: 1; background: #1e2a3a; color: #e0e0e0;
+    border: 1px solid #5d4037; border-radius: 8px;
+    padding: 8px 14px; font-size: 14px; outline: none;
+}
+.stage2-input:focus { border-color: #ffb74d; }
+.stage2-input::placeholder { color: #607d8b; }
+.stage2-controls { display: flex; flex-wrap: wrap; gap: 10px; align-items: center; }
+.stage2-controls label { color: #ffb74d; }
+.stage2-controls select, .stage2-controls input[type="number"] {
+    background: #1e2a3a; color: #e0e0e0;
+    border: 1px solid #5d4037; border-radius: 6px;
+    padding: 5px 8px; font-size: 12px; outline: none;
+}
+.stage2-controls select:focus, .stage2-controls input:focus { border-color: #ffb74d; }
+
+/* ── Toggle Button ──────────────────────────────────────────── */
+.toggle-btn {
+    background: transparent; color: #ffb74d; border: 1px solid #5d4037;
+    border-radius: 14px; padding: 2px 12px; font-size: 11px; cursor: pointer;
+    font-weight: 600; transition: all 0.15s;
+}
+.toggle-btn:hover { background: rgba(255,152,0,0.1); }
+.toggle-btn.active { background: rgba(255,152,0,0.15); border-color: #ffb74d; }
+
+/* ── Audio Upload Area ──────────────────────────────────────── */
+.upload-row {
+    display: none; gap: 10px; margin-bottom: 10px; align-items: center;
+}
+.upload-row.active { display: flex; }
+.upload-dropzone {
+    flex: 1; background: #1e2a3a; border: 2px dashed #37474f;
+    border-radius: 8px; padding: 12px 16px; cursor: pointer;
+    transition: all 0.2s; display: flex; align-items: center; gap: 10px;
+    min-height: 48px;
+}
+.upload-dropzone:hover, .upload-dropzone.dragover { border-color: #26c6da; background: #1a2e3a; }
+.upload-dropzone.has-file { border-color: #26c6da; border-style: solid; }
+.upload-dropzone .upload-icon { font-size: 20px; color: #546e7a; }
+.upload-dropzone .upload-text { font-size: 14px; color: #78909c; }
+.upload-dropzone .upload-filename { font-size: 14px; color: #26c6da; font-weight: 600; }
+.upload-dropzone input[type="file"] { display: none; }
+.upload-clear {
+    background: transparent; color: #ef5350; border: 1px solid #b71c1c;
+    border-radius: 6px; padding: 4px 10px; font-size: 11px; cursor: pointer;
+    font-weight: 600; display: none;
+}
+.upload-clear.visible { display: inline-block; }
+
+/* ── Music Similarity Reference ─────────────────────────────── */
+.sim-reference {
+    display: none; margin-bottom: 10px; padding: 8px 14px;
+    background: rgba(0,188,212,0.08); border: 1px solid #00838f;
+    border-radius: 8px; font-size: 13px; color: #80deea;
+    align-items: center; gap: 8px;
+}
+.sim-reference.active { display: flex; }
+.sim-reference .ref-label { font-weight: 600; color: #26c6da; }
+.sim-reference .ref-title { flex: 1; overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
+.sim-reference .ref-clear {
+    background: transparent; color: #ef5350; border: 1px solid #b71c1c;
+    border-radius: 6px; padding: 2px 8px; font-size: 10px; cursor: pointer;
+    font-weight: 600;
+}
+
+/* ── Find Similar Button ────────────────────────────────────── */
+.btn-find-similar {
+    background: rgba(0,188,212,0.1); color: #26c6da; border: 1px solid #00838f;
+    border-radius: 6px; padding: 3px 10px; font-size: 11px; cursor: pointer;
+    font-weight: 600; transition: all 0.15s; white-space: nowrap;
+}
+.btn-find-similar:hover { background: rgba(0,188,212,0.2); border-color: #26c6da; }
+
+/* ── Stats Bar ───────────────────────────────────────────────── */
+.stats-bar {
+    padding: 6px 24px; background: #141a22;
+    border-bottom: 1px solid #1e2a3a;
+    font-size: 12px; color: #607d8b;
+    display: flex; gap: 20px; flex-wrap: wrap; align-items: center;
+}
+.stats-bar .val { color: #4fc3f7; font-weight: 600; }
+.btn-download {
+    margin-left: auto;
+    background: transparent; color: #66bb6a; border: 1px solid #2e7d32;
+    border-radius: 6px; padding: 3px 12px; font-size: 11px;
+    font-weight: 600; cursor: pointer; transition: all 0.15s;
+    display: none;
+}
+.btn-download:hover { background: rgba(76,175,80,0.15); }
+
+/* ── Results ─────────────────────────────────────────────────── */
+.results-container { max-width: 1200px; margin: 0 auto; padding: 16px 24px; }
+.result-card {
+    background: linear-gradient(135deg, #1a2332 0%, #1e2a3a 100%);
+    border: 1px solid #263238; border-radius: 10px;
+    padding: 16px; margin-bottom: 12px;
+    transition: border-color 0.2s;
+}
+.result-card:hover { border-color: #37474f; }
+.card-top { display: flex; gap: 16px; align-items: flex-start; }
+.card-rank { font-size: 20px; font-weight: 700; color: #37474f; min-width: 32px; text-align: right; padding-top: 2px; }
+.card-main { flex: 1; min-width: 0; }
+.card-title {
+    font-size: 16px; font-weight: 700; color: #e7e9ea; margin-bottom: 4px;
+    overflow: hidden; text-overflow: ellipsis; white-space: nowrap;
+}
+
+/* ── Badges ──────────────────────────────────────────────────── */
+.badge-row { display: flex; flex-wrap: wrap; gap: 4px; margin: 6px 0; }
+.badge { display: inline-block; font-size: 10px; font-weight: 600; padding: 2px 8px; border-radius: 10px; }
+.badge-suno { background: rgba(33,150,243,0.15); color: #90caf9; border: 1px solid #2196f3; }
+.badge-udio { background: rgba(156,39,176,0.15); color: #ce93d8; border: 1px solid #9c27b0; }
+.badge-mureka { background: rgba(76,175,80,0.15); color: #a5d6a7; border: 1px solid #4caf50; }
+.badge-riffusion { background: rgba(255,152,0,0.15); color: #ffcc80; border: 1px solid #ff9800; }
+.badge-sonauto { background: rgba(0,188,212,0.15); color: #80deea; border: 1px solid #00bcd4; }
+.badge-genre { background: rgba(255,255,255,0.05); color: #b0bec5; border: 1px solid #455a64; }
+.badge-mood { background: rgba(233,30,99,0.1); color: #f48fb1; border: 1px solid #e91e63; }
+.badge-lang { background: rgba(103,58,183,0.15); color: #b39ddb; border: 1px solid #7e57c2; }
+.badge-instrumental { background: rgba(0,150,136,0.15); color: #80cbc4; border: 1px solid #009688; }
+.badge-vocal { background: rgba(255,87,34,0.1); color: #ffab91; border: 1px solid #ff5722; }
+.badge-nsfw-warn { background: rgba(255,152,0,0.2); color: #ffcc80; border: 1px solid #ff9800; }
+.badge-nsfw-danger { background: rgba(244,67,54,0.2); color: #ef9a9a; border: 1px solid #f44336; }
+
+/* ── Metrics Row ─────────────────────────────────────────────── */
+.metrics-row { display: flex; flex-wrap: wrap; gap: 12px; margin: 8px 0; font-size: 12px; color: #90a4ae; }
+.metric-item { display: flex; align-items: center; gap: 3px; }
+.metric-label { color: #607d8b; }
+.metric-val { font-weight: 700; }
+
+/* ── Aesthetics Bar ──────────────────────────────────────────── */
+.aesthetics-bar { height: 4px; border-radius: 2px; margin: 4px 0; background: #263238; position: relative; max-width: 100px; }
+.aesthetics-bar-fill { height: 100%; border-radius: 2px; position: absolute; left: 0; top: 0; }
+
+/* ── Caption (full display) ──────────────────────────────────── */
+.card-caption {
+    font-size: 12px; color: #90a4ae; margin: 8px 0;
+    line-height: 1.6; background: rgba(0,0,0,0.15);
+    padding: 10px 12px; border-radius: 6px; border-left: 3px solid #37474f;
+}
+
+/* ── Audio Player ────────────────────────────────────────────── */
+.card-audio { margin-top: 8px; display: flex; align-items: center; gap: 8px; }
+.card-audio audio { flex: 1; height: 32px; border-radius: 6px; }
+
+/* ── Score Display ───────────────────────────────────────────── */
+.card-score { text-align: right; min-width: 80px; }
+.card-score .score-val { font-size: 22px; font-weight: 800; color: #4fc3f7; }
+.card-score .score-label { font-size: 10px; color: #607d8b; }
+.card-scores-dual { text-align: right; min-width: 90px; }
+.card-scores-dual .s1 { font-size: 11px; color: #607d8b; }
+.card-scores-dual .s1 .v { color: #64b5f6; font-weight: 700; }
+.card-scores-dual .s2 { font-size: 16px; margin-top: 2px; }
+.card-scores-dual .s2 .v { color: #ffb74d; font-weight: 800; }
+.card-scores-dual .s2 .l { font-size: 10px; color: #8d6e63; }
+
+/* ── Tags display ────────────────────────────────────────────── */
+.tags-text { font-size: 11px; color: #78909c; overflow: hidden; text-overflow: ellipsis; white-space: nowrap; max-width: 600px; }
+
+/* ── Loading / Empty ─────────────────────────────────────────── */
+.loading { text-align: center; padding: 60px; color: #546e7a; font-size: 16px; }
+.loading .spinner {
+    display: inline-block; width: 30px; height: 30px;
+    border: 3px solid #263238; border-top-color: #4fc3f7;
+    border-radius: 50%; animation: spin 0.8s linear infinite; margin-bottom: 12px;
+}
+@keyframes spin { to { transform: rotate(360deg); } }
+.empty-state { text-align: center; padding: 80px 24px; color: #546e7a; }
+.empty-state .icon { font-size: 48px; margin-bottom: 16px; }
+.empty-state p { font-size: 14px; max-width: 400px; margin: 0 auto; }
+</style>
+</head>
+<body>
+
+<div class="header">
+    <h1>LAION-Tunes <span>Search</span></h1>
+
+    <!-- Text query row (hidden in music_similarity mode) -->
+    <div class="search-row" id="query-row">
+        <input type="text" class="search-input" id="query" placeholder="Stage 1: Search music... e.g. 'dreamy ambient synth pad' or 'aggressive metal guitar solo'" autofocus>
+        <button class="btn-search" id="btn-search-main" onclick="doSearch()">Search</button>
+    </div>
+
+    <!-- Audio upload row (shown in music_similarity mode) -->
+    <div class="upload-row" id="upload-row">
+        <div class="upload-dropzone" id="upload-dropzone" onclick="document.getElementById('audio-file-input').click()">
+            <span class="upload-icon">&#9835;</span>
+            <span class="upload-text" id="upload-text">Drop audio file here or click to browse (first 30s used)</span>
+            <span class="upload-filename" id="upload-filename" style="display:none"></span>
+            <input type="file" id="audio-file-input" accept="audio/*,.mp3,.wav,.flac,.ogg,.m4a,.mp4,.webm,.opus">
+        </div>
+        <button class="upload-clear" id="upload-clear" onclick="clearUpload()">Clear</button>
+        <button class="btn-search" onclick="doSearch()">Search</button>
+    </div>
+
+    <!-- Music similarity reference indicator -->
+    <div class="sim-reference" id="sim-reference">
+        <span class="ref-label">Similar to:</span>
+        <span class="ref-title" id="sim-ref-title"></span>
+        <button class="ref-clear" onclick="clearSimilarityRef()">Clear</button>
+    </div>
+
+    <!-- Audio preview player for uploaded file -->
+    <div class="sim-reference" id="upload-preview" style="display:none; background:rgba(0,150,136,0.08); border-color:#00695c;">
+        <span class="ref-label" style="color:#26a69a;">Uploaded:</span>
+        <span class="ref-title" id="upload-preview-name" style="color:#80cbc4;"></span>
+        <audio controls id="upload-audio-player" style="height:28px; max-width:300px; flex-shrink:0;"></audio>
+    </div>
+
+    <div class="search-row" id="negative-row" style="display:none">
+        <input type="text" class="search-input" id="negative_query" placeholder="Negative prompt: exclude results matching this... e.g. 'vocals singing lyrics'" style="border-color:#ef5350; color:#ef9a9a;">
+        <div class="control-group" style="min-width:100px">
+            <label style="color:#ef5350">Weight:</label>
+            <input type="range" id="neg_weight" min="0.1" max="1.0" step="0.1" value="0.7" style="accent-color:#ef5350;" oninput="document.getElementById('neg-val').textContent=this.value">
+            <span class="aes-val" id="neg-val" style="color:#ef5350">0.7</span>
+        </div>
+    </div>
+
+    <!-- Row 1: Main controls -->
+    <div class="controls">
+        <div class="control-group">
+            <label>Mode:</label>
+            <select id="search_type" onchange="updateControls()">
+                <option value="bm25" selected>BM25 Text</option>
+                <option value="vector">Vector Similarity</option>
+                <option value="combined">Combined</option>
+                <option value="music_similarity">Music Similarity</option>
+            </select>
+        </div>
+        <div class="control-group" id="vector-field-group" style="display:none">
+            <label>Vector:</label>
+            <select id="vector_field">
+                <option value="caption" selected>Whisper Caption</option>
+                <option value="tag">Tags</option>
+                <option value="lyric">Lyrics</option>
+                <option value="mood">Mood</option>
+                <option value="transcription">Transcription</option>
+            </select>
+        </div>
+        <div class="control-group" id="bm25-field-group">
+            <label>BM25 Field:</label>
+            <select id="bm25_field">
+                <option value="caption" selected>Whisper Caption</option>
+                <option value="tags">Tags</option>
+                <option value="transcription">Transcription</option>
+                <option value="lyrics_hashed">Lyrics (hashed)</option>
+            </select>
+        </div>
+        <div class="control-group">
+            <label>Rank by:</label>
+            <select id="rank_by">
+                <option value="similarity" selected>Relevance</option>
+                <option value="music_similarity">Music Similarity</option>
+                <option value="aesthetics">Aesthetics Score</option>
+                <option value="plays">Play Count</option>
+                <option value="likes">Like Count</option>
+            </select>
+        </div>
+        <div class="control-group">
+            <label>K:</label>
+            <select id="top_k" onchange="syncTopKInput(this)">
+                <option value="20">20</option>
+                <option value="50" selected>50</option>
+                <option value="100">100</option>
+                <option value="200">200</option>
+                <option value="500">500</option>
+                <option value="1000">1,000</option>
+                <option value="5000">5,000</option>
+                <option value="10000">10,000</option>
+                <option value="custom">Custom...</option>
+            </select>
+            <input type="number" id="top_k_custom" min="1" max="50000" value="50" style="width:65px; display:none">
+        </div>
+        <button class="toggle-btn" id="stage2-toggle" onclick="toggleStage2()">+ Stage 2</button>
+    </div>
+
+    <!-- Stage 2 Panel -->
+    <div class="stage2-panel" id="stage2-panel">
+        <div class="stage2-header">
+            <span class="label">Stage 2 Refinement</span>
+            <span style="font-size:11px; color:#8d6e63;">Re-filter Stage 1 results by a second query</span>
+        </div>
+        <div class="stage2-row">
+            <input type="text" class="stage2-input" id="s2_query" placeholder="Stage 2 query... e.g. 'angry little dog' or 'hip-hop beat'">
+        </div>
+        <div class="stage2-controls">
+            <div class="control-group">
+                <label>Mode:</label>
+                <select id="s2_search_type" onchange="updateS2Controls()">
+                    <option value="vector" selected>Vector Similarity</option>
+                    <option value="bm25">BM25 Text</option>
+                </select>
+            </div>
+            <div class="control-group" id="s2-vector-field-group">
+                <label>Vector:</label>
+                <select id="s2_vector_field">
+                    <option value="caption" selected>Whisper Caption</option>
+                    <option value="tag">Tags</option>
+                    <option value="lyric">Lyrics</option>
+                    <option value="mood">Mood</option>
+                    <option value="transcription">Transcription</option>
+                </select>
+            </div>
+            <div class="control-group" id="s2-bm25-field-group" style="display:none">
+                <label>BM25 Field:</label>
+                <select id="s2_bm25_field">
+                    <option value="caption" selected>Whisper Caption</option>
+                    <option value="tags">Tags</option>
+                    <option value="transcription">Transcription</option>
+                    <option value="lyrics_hashed">Lyrics (hashed)</option>
+                </select>
+            </div>
+            <div class="control-group">
+                <label>Stage 2 K:</label>
+                <select id="s2_top_k" onchange="syncS2TopKInput(this)">
+                    <option value="20">20</option>
+                    <option value="50" selected>50</option>
+                    <option value="100">100</option>
+                    <option value="200">200</option>
+                    <option value="custom">Custom...</option>
+                </select>
+                <input type="number" id="s2_top_k_custom" min="1" max="50000" value="50" style="width:65px; display:none">
+            </div>
+        </div>
+    </div>
+
+    <!-- Row 2: Filters -->
+    <div class="filters">
+        <div class="filter-section">
+            <span class="filter-section-label">Vocals:</span>
+            <span class="chip active" onclick="setVocal(this, null)">All</span>
+            <span class="chip" onclick="setVocal(this, 'instrumental')">Instrumental</span>
+            <span class="chip" onclick="setVocal(this, 'vocals')">With Vocals</span>
+        </div>
+        <div class="filter-section">
+            <span class="filter-section-label">Min Duration:</span>
+            <div class="control-group">
+                <input type="number" id="min_duration" value="60" min="0" max="600" step="10" style="width:55px">
+                <label>sec</label>
+            </div>
+        </div>
+        <div class="filter-section">
+            <span class="filter-section-label">Min Aesthetics:</span>
+            <div class="control-group">
+                <input type="range" id="min_aesthetics" min="0" max="5" step="0.1" value="0" oninput="document.getElementById('aes-val').textContent=this.value">
+                <span class="aes-val" id="aes-val">0</span>
+            </div>
+        </div>
+        <div class="filter-section">
+            <span class="filter-section-label">Subset:</span>
+            <span class="chip active" onclick="setSubset(this, null)">All</span>
+            <span class="chip" onclick="setSubset(this, 'suno')">Suno</span>
+            <span class="chip" onclick="setSubset(this, 'udio')">Udio</span>
+            <span class="chip" onclick="setSubset(this, 'mureka')">Mureka</span>
+            <span class="chip" onclick="setSubset(this, 'riffusion')">Riffusion</span>
+            <span class="chip" onclick="setSubset(this, 'sonauto')">Sonauto</span>
+        </div>
+        <div class="filter-section">
+            <span class="filter-section-label">Safety:</span>
+            <span class="chip active" onclick="setNsfw(this, null)">All</span>
+            <span class="chip" onclick="setNsfw(this, 'sfw_only')">SFW Only</span>
+            <span class="chip" onclick="setNsfw(this, 'nsfw_only')">NSFW Only</span>
+        </div>
+    </div>
+
+    <!-- Row 3: Language filters -->
+    <div class="filters">
+        <div class="filter-section">
+            <span class="filter-section-label">Languages:</span>
+            <span class="chip active" onclick="setAllLangs(this)">All</span>
+            <div class="lang-panel" id="lang-panel">
+                <!-- Populated dynamically from /api/stats -->
+            </div>
+        </div>
+    </div>
+</div>
+
+<div class="stats-bar" id="stats-bar">
+    <span>Total: <span class="val" id="stat-total">--</span></span>
+    <span>Results: <span class="val" id="stat-results">--</span></span>
+    <span>Filtered: <span class="val" id="stat-filtered">--</span></span>
+    <span id="stat-stage2-wrap" style="display:none">Stage2: <span class="val" id="stat-stage2">--</span></span>
+    <span id="stat-whisper-wrap" style="display:none">Whisper Index: <span class="val" id="stat-whisper">--</span></span>
+    <span id="stat-cache-wrap" style="display:none"><span class="val" id="stat-cache" style="color:#26a69a;">Cached</span></span>
+    <span>Time: <span class="val" id="stat-time">--</span></span>
+    <span>Embedding: <span class="val" id="stat-emb">--</span></span>
+    <button class="btn-download" id="btn-download" onclick="downloadResults()">Download JSON</button>
+</div>
+
+<div class="results-container" id="results">
+    <div class="empty-state">
+        <div class="icon">&#9835;</div>
+        <p>Search the LAION-Tunes dataset of 908K AI-generated music tracks. Use BM25 for text matching, Vector for semantic similarity, or Music Similarity for audio-based search. Enable Stage 2 to refine results with a second query.</p>
+    </div>
+</div>
+
+<script>
+// ── State ───────────────────────────────────────────────────────
+let selectedSubset = null;
+let selectedVocal = null;     // null | "instrumental" | "vocals"
+let selectedNsfw = null;      // null | "sfw_only" | "nsfw_only"
+let selectedLangs = null;     // null = all, or Set of codes
+let allLanguages = {};        // code -> count
+let isSearching = false;
+let lastResults = null;       // store for download
+let stage2Active = false;
+let whisperEmbCount = 0;
+
+// Music similarity state
+let uploadedAudioFile = null;
+let similarityRowId = null;
+let similarityRefTitle = null;
+
+// Language code to display name
+const LANG_NAMES = {
+    en:'English', es:'Spanish', fr:'French', de:'German', pt:'Portuguese',
+    it:'Italian', nl:'Dutch', ru:'Russian', ja:'Japanese', ko:'Korean',
+    zh:'Chinese', ar:'Arabic', hi:'Hindi', tr:'Turkish', pl:'Polish',
+    sv:'Swedish', da:'Danish', no:'Norwegian', fi:'Finnish', cs:'Czech',
+    ro:'Romanian', hu:'Hungarian', el:'Greek', th:'Thai', vi:'Vietnamese',
+    id:'Indonesian', ms:'Malay', tl:'Tagalog', uk:'Ukrainian', bg:'Bulgarian',
+    hr:'Croatian', sk:'Slovak', sl:'Slovenian', lt:'Lithuanian', lv:'Latvian',
+    et:'Estonian', ca:'Catalan', gl:'Galician', eu:'Basque', cy:'Welsh',
+    af:'Afrikaans', sw:'Swahili', sq:'Albanian', mk:'Macedonian', sr:'Serbian',
+    unknown:'Unknown'
+};
+
+// ── Stage 2 Toggle ──────────────────────────────────────────────
+function toggleStage2() {
+    stage2Active = !stage2Active;
+    const panel = document.getElementById('stage2-panel');
+    const btn = document.getElementById('stage2-toggle');
+    if (stage2Active) {
+        panel.classList.add('active');
+        btn.classList.add('active');
+        btn.textContent = '- Stage 2';
+    } else {
+        panel.classList.remove('active');
+        btn.classList.remove('active');
+        btn.textContent = '+ Stage 2';
+    }
+}
+
+function updateS2Controls() {
+    const mode = document.getElementById('s2_search_type').value;
+    document.getElementById('s2-vector-field-group').style.display = mode === 'vector' ? 'flex' : 'none';
+    document.getElementById('s2-bm25-field-group').style.display = mode === 'bm25' ? 'flex' : 'none';
+}
+
+// ── Custom K inputs ─────────────────────────────────────────────
+function syncTopKInput(sel) {
+    const custom = document.getElementById('top_k_custom');
+    if (sel.value === 'custom') {
+        custom.style.display = 'inline-block';
+        custom.focus();
+    } else {
+        custom.style.display = 'none';
+        custom.value = sel.value;
+    }
+}
+function syncS2TopKInput(sel) {
+    const custom = document.getElementById('s2_top_k_custom');
+    if (sel.value === 'custom') {
+        custom.style.display = 'inline-block';
+        custom.focus();
+    } else {
+        custom.style.display = 'none';
+        custom.value = sel.value;
+    }
+}
+function getTopK() {
+    const sel = document.getElementById('top_k');
+    if (sel.value === 'custom') return parseInt(document.getElementById('top_k_custom').value) || 50;
+    return parseInt(sel.value);
+}
+function getS2TopK() {
+    const sel = document.getElementById('s2_top_k');
+    if (sel.value === 'custom') return parseInt(document.getElementById('s2_top_k_custom').value) || 50;
+    return parseInt(sel.value);
+}
+
+// ── Controls visibility ─────────────────────────────────────────
+function updateControls() {
+    const mode = document.getElementById('search_type').value;
+    const isMusicSim = mode === 'music_similarity';
+    const isVector = mode === 'vector' || mode === 'combined';
+
+    // Show/hide query row vs upload row
+    document.getElementById('query-row').style.display = isMusicSim ? 'none' : 'flex';
+    document.getElementById('upload-row').classList.toggle('active', isMusicSim);
+
+    // Show/hide vector/BM25 field selectors
+    document.getElementById('vector-field-group').style.display = (isVector && !isMusicSim) ? 'flex' : 'none';
+    document.getElementById('bm25-field-group').style.display =
+        (mode === 'bm25' || mode === 'combined') ? 'flex' : 'none';
+
+    // Show/hide negative prompt row (only for vector/combined)
+    document.getElementById('negative-row').style.display = (isVector && !isMusicSim) ? 'flex' : 'none';
+
+    // Show stage2 toggle (always available, including music_similarity)
+    document.getElementById('stage2-toggle').style.display = '';
+
+    // Show/hide similarity reference
+    updateSimReference();
+
+    // Default rank_by to music_similarity when in that mode
+    if (isMusicSim) {
+        document.getElementById('rank_by').value = 'music_similarity';
+    } else if (document.getElementById('rank_by').value === 'music_similarity') {
+        document.getElementById('rank_by').value = 'similarity';
+    }
+}
+
+// ── Audio Upload ────────────────────────────────────────────────
+const dropzone = document.getElementById('upload-dropzone');
+const fileInput = document.getElementById('audio-file-input');
+
+fileInput.addEventListener('change', function() {
+    if (this.files.length > 0) {
+        setUploadFile(this.files[0]);
+    }
+});
+
+dropzone.addEventListener('dragover', function(e) {
+    e.preventDefault();
+    dropzone.classList.add('dragover');
+});
+dropzone.addEventListener('dragleave', function() {
+    dropzone.classList.remove('dragover');
+});
+dropzone.addEventListener('drop', function(e) {
+    e.preventDefault();
+    dropzone.classList.remove('dragover');
+    if (e.dataTransfer.files.length > 0) {
+        setUploadFile(e.dataTransfer.files[0]);
+    }
+});
+
+let uploadAudioObjectURL = null;
+
+function setUploadFile(file) {
+    uploadedAudioFile = file;
+    similarityRowId = null; // Clear any sample reference
+    similarityRefTitle = null;
+    document.getElementById('upload-text').style.display = 'none';
+    document.getElementById('upload-filename').style.display = 'inline';
+    document.getElementById('upload-filename').textContent = file.name + ' (' + (file.size / 1024 / 1024).toFixed(1) + ' MB)';
+    document.getElementById('upload-clear').classList.add('visible');
+    dropzone.classList.add('has-file');
+
+    // Audio preview player
+    if (uploadAudioObjectURL) URL.revokeObjectURL(uploadAudioObjectURL);
+    uploadAudioObjectURL = URL.createObjectURL(file);
+    const player = document.getElementById('upload-audio-player');
+    player.src = uploadAudioObjectURL;
+    document.getElementById('upload-preview-name').textContent = file.name;
+    document.getElementById('upload-preview').style.display = 'flex';
+
+    updateSimReference();
+}
+
+function clearUpload() {
+    uploadedAudioFile = null;
+    fileInput.value = '';
+    document.getElementById('upload-text').style.display = 'inline';
+    document.getElementById('upload-filename').style.display = 'none';
+    document.getElementById('upload-clear').classList.remove('visible');
+    dropzone.classList.remove('has-file');
+
+    // Clear audio preview player
+    if (uploadAudioObjectURL) {
+        URL.revokeObjectURL(uploadAudioObjectURL);
+        uploadAudioObjectURL = null;
+    }
+    document.getElementById('upload-audio-player').src = '';
+    document.getElementById('upload-preview').style.display = 'none';
+
+    updateSimReference();
+}
+
+// ── Similarity Reference ────────────────────────────────────────
+function updateSimReference() {
+    const ref = document.getElementById('sim-reference');
+    const mode = document.getElementById('search_type').value;
+    if (mode === 'music_similarity' && similarityRowId && similarityRefTitle) {
+        ref.classList.add('active');
+        document.getElementById('sim-ref-title').textContent = similarityRefTitle;
+    } else {
+        ref.classList.remove('active');
+    }
+}
+
+function clearSimilarityRef() {
+    similarityRowId = null;
+    similarityRefTitle = null;
+    updateSimReference();
+}
+
+function findSimilar(rowId, title) {
+    // Switch to music similarity mode
+    document.getElementById('search_type').value = 'music_similarity';
+    similarityRowId = rowId;
+    similarityRefTitle = title;
+    uploadedAudioFile = null;
+    fileInput.value = '';
+    document.getElementById('upload-text').style.display = 'inline';
+    document.getElementById('upload-filename').style.display = 'none';
+    document.getElementById('upload-clear').classList.remove('visible');
+    dropzone.classList.remove('has-file');
+    // Clear audio preview
+    if (uploadAudioObjectURL) {
+        URL.revokeObjectURL(uploadAudioObjectURL);
+        uploadAudioObjectURL = null;
+    }
+    document.getElementById('upload-audio-player').src = '';
+    document.getElementById('upload-preview').style.display = 'none';
+    updateControls();
+    doSearch();
+}
+
+// ── Chip group helpers ──────────────────────────────────────────
+function setVocal(el, val) {
+    el.parentElement.querySelectorAll('.chip').forEach(c => c.classList.remove('active'));
+    el.classList.add('active');
+    selectedVocal = val;
+}
+
+function setSubset(el, val) {
+    el.parentElement.querySelectorAll('.chip').forEach(c => c.classList.remove('active'));
+    el.classList.add('active');
+    selectedSubset = val;
+}
+
+function setNsfw(el, val) {
+    el.parentElement.querySelectorAll('.chip').forEach(c => c.classList.remove('active'));
+    el.classList.add('active');
+    selectedNsfw = val;
+}
+
+function setAllLangs(el) {
+    selectedLangs = null;
+    el.classList.add('active');
+    document.querySelectorAll('.lang-chip').forEach(c => c.classList.remove('active'));
+}
+
+function toggleLang(el, code) {
+    // Deactivate "All" chip
+    const allChip = el.closest('.filter-section').querySelector('.chip');
+    allChip.classList.remove('active');
+
+    el.classList.toggle('active');
+    // Collect active languages
+    const active = document.querySelectorAll('.lang-chip.active');
+    if (active.length === 0) {
+        selectedLangs = null;
+        allChip.classList.add('active');
+    } else {
+        selectedLangs = new Set();
+        active.forEach(c => selectedLangs.add(c.dataset.lang));
+    }
+}
+
+// ── Download Results as JSON ────────────────────────────────────
+function downloadResults() {
+    if (!lastResults || !lastResults.results || lastResults.results.length === 0) return;
+    const blob = new Blob([JSON.stringify(lastResults.results, null, 2)], { type: 'application/json' });
+    const url = URL.createObjectURL(blob);
+    const a = document.createElement('a');
+    a.href = url;
+    a.download = 'laion-tunes-results.json';
+    a.click();
+    URL.revokeObjectURL(url);
+}
+
+// ── Search ──────────────────────────────────────────────────────
+async function doSearch() {
+    if (isSearching) return;
+
+    const mode = document.getElementById('search_type').value;
+    const resultsDiv = document.getElementById('results');
+
+    // Music similarity mode
+    if (mode === 'music_similarity') {
+        return doMusicSimilaritySearch();
+    }
+
+    const query = document.getElementById('query').value.trim();
+    if (!query) return;
+
+    isSearching = true;
+    resultsDiv.innerHTML = '<div class="loading"><div class="spinner"></div><br>Searching...</div>';
+
+    const minAes = parseFloat(document.getElementById('min_aesthetics').value);
+    const minDur = parseFloat(document.getElementById('min_duration').value);
+
+    const negQuery = document.getElementById('negative_query').value.trim();
+    const negWeight = parseFloat(document.getElementById('neg_weight').value);
+
+    const body = {
+        query: query,
+        negative_query: negQuery || null,
+        search_type: document.getElementById('search_type').value,
+        vector_field: document.getElementById('vector_field').value,
+        bm25_field: document.getElementById('bm25_field').value,
+        rank_by: document.getElementById('rank_by').value,
+        min_aesthetics: minAes > 0 ? minAes : null,
+        min_similarity: null,
+        subset_filter: selectedSubset,
+        vocal_filter: selectedVocal,
+        min_duration: minDur > 0 ? minDur : null,
+        languages: selectedLangs ? Array.from(selectedLangs) : null,
+        negative_weight: negWeight,
+        nsfw_filter: selectedNsfw,
+        top_k: getTopK(),
+    };
+
+    // Stage 2
+    if (stage2Active) {
+        const s2q = document.getElementById('s2_query').value.trim();
+        if (s2q) {
+            body.stage2_enabled = true;
+            body.stage2_query = s2q;
+            body.stage2_search_type = document.getElementById('s2_search_type').value;
+            body.stage2_vector_field = document.getElementById('s2_vector_field').value;
+            body.stage2_bm25_field = document.getElementById('s2_bm25_field').value;
+            body.stage2_top_k = getS2TopK();
+        }
+    }
+
+    try {
+        const resp = await fetch('/api/search', {
+            method: 'POST',
+            headers: { 'Content-Type': 'application/json' },
+            body: JSON.stringify(body),
+        });
+        const data = await resp.json();
+        lastResults = data;
+        renderResults(data);
+    } catch (err) {
+        resultsDiv.innerHTML = `<div class="empty-state"><p>Error: ${err.message}</p></div>`;
+    } finally {
+        isSearching = false;
+    }
+}
+
+async function doMusicSimilaritySearch() {
+    const resultsDiv = document.getElementById('results');
+
+    // Decide: upload audio or use pre-computed embedding
+    if (similarityRowId) {
+        // Use pre-computed embedding
+        isSearching = true;
+        resultsDiv.innerHTML = '<div class="loading"><div class="spinner"></div><br>Finding similar tracks...</div>';
+
+        const minAes = parseFloat(document.getElementById('min_aesthetics').value);
+        const minDur = parseFloat(document.getElementById('min_duration').value);
+
+        const body = {
+            row_id: similarityRowId,
+            top_k: getTopK(),
+            rank_by: document.getElementById('rank_by').value,
+            min_aesthetics: minAes > 0 ? minAes : null,
+            subset_filter: selectedSubset,
+            vocal_filter: selectedVocal,
+            min_duration: minDur > 0 ? minDur : null,
+            languages: selectedLangs ? Array.from(selectedLangs) : null,
+            nsfw_filter: selectedNsfw,
+        };
+
+        // Stage 2 for similar search
+        if (stage2Active) {
+            const s2q = document.getElementById('s2_query').value.trim();
+            if (s2q) {
+                body.stage2_enabled = true;
+                body.stage2_query = s2q;
+                body.stage2_search_type = document.getElementById('s2_search_type').value;
+                body.stage2_vector_field = document.getElementById('s2_vector_field').value;
+                body.stage2_bm25_field = document.getElementById('s2_bm25_field').value;
+                body.stage2_top_k = getS2TopK();
+            }
+        }
+
+        try {
+            const resp = await fetch('/api/search_similar', {
+                method: 'POST',
+                headers: { 'Content-Type': 'application/json' },
+                body: JSON.stringify(body),
+            });
+            const data = await resp.json();
+            if (data.error) {
+                resultsDiv.innerHTML = `<div class="empty-state"><p>Error: ${data.error}</p></div>`;
+            } else {
+                lastResults = data;
+                renderResults(data);
+            }
+        } catch (err) {
+            resultsDiv.innerHTML = `<div class="empty-state"><p>Error: ${err.message}</p></div>`;
+        } finally {
+            isSearching = false;
+        }
+
+    } else if (uploadedAudioFile) {
+        // Upload audio
+        isSearching = true;
+        resultsDiv.innerHTML = '<div class="loading"><div class="spinner"></div><br>Processing audio & searching...</div>';
+
+        const minAes = parseFloat(document.getElementById('min_aesthetics').value);
+        const minDur = parseFloat(document.getElementById('min_duration').value);
+
+        const formData = new FormData();
+        formData.append('audio', uploadedAudioFile);
+        formData.append('top_k', getTopK());
+        formData.append('rank_by', document.getElementById('rank_by').value);
+        formData.append('subset_filter', selectedSubset || '');
+        formData.append('vocal_filter', selectedVocal || '');
+        formData.append('min_duration', minDur > 0 ? minDur : '');
+        formData.append('min_aesthetics', minAes > 0 ? minAes : '');
+        formData.append('languages', selectedLangs ? Array.from(selectedLangs).join(',') : '');
+        formData.append('nsfw_filter', selectedNsfw || '');
+
+        // Stage 2 for audio upload search
+        if (stage2Active) {
+            const s2q = document.getElementById('s2_query').value.trim();
+            if (s2q) {
+                formData.append('stage2_enabled', 'true');
+                formData.append('stage2_query', s2q);
+                formData.append('stage2_search_type', document.getElementById('s2_search_type').value);
+                formData.append('stage2_vector_field', document.getElementById('s2_vector_field').value);
+                formData.append('stage2_bm25_field', document.getElementById('s2_bm25_field').value);
+                formData.append('stage2_top_k', getS2TopK());
+            }
+        }
+
+        try {
+            const resp = await fetch('/api/search_by_audio', {
+                method: 'POST',
+                body: formData,
+            });
+            const data = await resp.json();
+            if (data.error) {
+                resultsDiv.innerHTML = `<div class="empty-state"><p>Error: ${data.error}</p></div>`;
+            } else {
+                lastResults = data;
+                renderResults(data);
+            }
+        } catch (err) {
+            resultsDiv.innerHTML = `<div class="empty-state"><p>Error: ${err.message}</p></div>`;
+        } finally {
+            isSearching = false;
+        }
+
+    } else {
+        resultsDiv.innerHTML = '<div class="empty-state"><p>Upload an audio file or click "Find Similar" on a result to search by music similarity.</p></div>';
+    }
+}
+
+// ── Render Results ──────────────────────────────────────────────
+function renderResults(data) {
+    document.getElementById('stat-total').textContent = data.total_tracks?.toLocaleString() || '--';
+    document.getElementById('stat-results').textContent = data.results?.length || 0;
+    document.getElementById('stat-filtered').textContent = data.total_filtered?.toLocaleString() || '--';
+    document.getElementById('stat-time').textContent = data.search_time_ms ? data.search_time_ms.toFixed(0) + 'ms' : '--';
+    document.getElementById('stat-emb').textContent = data.query_embedding_time_ms ? data.query_embedding_time_ms.toFixed(0) + 'ms' : '--';
+
+    // Stage 2 stats
+    const s2wrap = document.getElementById('stat-stage2-wrap');
+    if (data.stage2) {
+        s2wrap.style.display = '';
+        document.getElementById('stat-stage2').textContent = data.stage2.matched + ' matched, ' + data.stage2.returned + ' returned';
+    } else {
+        s2wrap.style.display = 'none';
+    }
+
+    // Whisper stats
+    const wWrap = document.getElementById('stat-whisper-wrap');
+    if (data.search_type === 'music_similarity') {
+        wWrap.style.display = '';
+        document.getElementById('stat-whisper').textContent = (whisperEmbCount || 0).toLocaleString();
+    } else {
+        wWrap.style.display = 'none';
+    }
+
+    // Cache hit indicator
+    const cWrap = document.getElementById('stat-cache-wrap');
+    cWrap.style.display = data.cache_hit ? '' : 'none';
+
+    // Download button
+    const dlBtn = document.getElementById('btn-download');
+    dlBtn.style.display = (data.results && data.results.length > 0) ? 'inline-block' : 'none';
+
+    const resultsDiv = document.getElementById('results');
+    if (!data.results || data.results.length === 0) {
+        resultsDiv.innerHTML = '<div class="empty-state"><p>No results found. Try a different query or broaden your filters.</p></div>';
+        return;
+    }
+
+    const hasStage2 = !!data.stage2;
+    const isMusicSim = data.search_type === 'music_similarity';
+    let html = '';
+    data.results.forEach((r, idx) => {
+        const subsetClass = 'badge-' + (r.subset || 'unknown');
+
+        // Genre badges (max 5)
+        let genreBadges = '';
+        (r.genre_tags || []).slice(0, 5).forEach(g => {
+            genreBadges += `<span class="badge badge-genre">${esc(g)}</span>`;
+        });
+
+        // Mood badges
+        let moodBadges = '';
+        if (r.mood_text) {
+            r.mood_text.split(',').slice(0, 4).forEach(m => {
+                m = m.trim();
+                if (m) moodBadges += `<span class="badge badge-mood">${esc(m)}</span>`;
+            });
+        } else if (r.emotion_tags && r.emotion_tags.length) {
+            r.emotion_tags.slice(0, 4).forEach(e => {
+                moodBadges += `<span class="badge badge-mood">${esc(e)}</span>`;
+            });
+        }
+
+        // Language badge
+        const langCode = r.language || 'unknown';
+        const langName = LANG_NAMES[langCode] || langCode;
+        const langBadge = `<span class="badge badge-lang">${esc(langName)}</span>`;
+
+        // Instrumental / Vocal badge
+        const vocalBadge = r.is_instrumental
+            ? '<span class="badge badge-instrumental">Instrumental</span>'
+            : '<span class="badge badge-vocal">Vocals</span>';
+
+        // NSFW safety badge
+        let nsfwBadge = '';
+        if (r.nsfw_overall_label === 'very_likely_nsfw') {
+            const parts = [];
+            if (r.nsfw_gore_label !== 'likely_sfw') parts.push('Gore');
+            if (r.nsfw_sexual_label !== 'likely_sfw') parts.push('Sexual');
+            if (r.nsfw_hate_label !== 'likely_sfw') parts.push('Hate');
+            nsfwBadge = `<span class="badge badge-nsfw-danger" title="Gore:${r.nsfw_gore_sim||'?'} Sexual:${r.nsfw_sexual_sim||'?'} Hate:${r.nsfw_hate_sim||'?'}">NSFW: ${parts.join(', ')}</span>`;
+        } else if (r.nsfw_overall_label === 'likely_nsfw') {
+            const parts = [];
+            if (r.nsfw_gore_label !== 'likely_sfw') parts.push('Gore');
+            if (r.nsfw_sexual_label !== 'likely_sfw') parts.push('Sexual');
+            if (r.nsfw_hate_label !== 'likely_sfw') parts.push('Hate');
+            nsfwBadge = `<span class="badge badge-nsfw-warn" title="Gore:${r.nsfw_gore_sim||'?'} Sexual:${r.nsfw_sexual_sim||'?'} Hate:${r.nsfw_hate_sim||'?'}">Caution: ${parts.join(', ')}</span>`;
+        }
+
+        // Aesthetics
+        const aes = r.score_average || 0;
+        const aesPct = Math.min(100, (aes / 5) * 100);
+        const aesColor = aes >= 3.5 ? '#66bb6a' : aes >= 2.5 ? '#ffa726' : '#ef5350';
+
+        // Score display
+        let scoreDisplay = '';
+        if (hasStage2 && r.stage2_score !== undefined) {
+            // Dual score: show stage1 small + stage2 prominent
+            const s1v = r.stage1_score !== undefined ? (r.score_type === 'bm25' ? r.stage1_score.toFixed(2) : (r.stage1_score*100).toFixed(1)+'%') : '--';
+            const s2v = r.score_type === 'bm25' ? r.stage2_score.toFixed(2) : (r.stage2_score*100).toFixed(1)+'%';
+            const s2label = r.score_type === 'bm25' ? 'BM25' : 'Sim';
+            scoreDisplay = `<div class="card-scores-dual">
+                <div class="s1">S1: <span class="v">${s1v}</span></div>
+                <div class="s2"><span class="v">${s2v}</span> <span class="l">${s2label}</span></div>
+            </div>`;
+        } else if (r.score !== null && r.score !== undefined) {
+            let label = '', val = r.score;
+            switch (r.score_type) {
+                case 'cosine_similarity': label = isMusicSim ? 'Audio Sim' : 'Similarity'; val = (val * 100).toFixed(1) + '%'; break;
+                case 'bm25': label = 'BM25'; val = val.toFixed(2); break;
+                case 'aesthetics': label = 'Aesthetics'; val = val.toFixed(2); break;
+                case 'play_count': label = 'Plays'; val = Number(val).toLocaleString(); break;
+                case 'upvote_count': label = 'Likes'; val = Number(val).toLocaleString(); break;
+                default: label = r.score_type; val = val.toFixed(2);
+            }
+            scoreDisplay = `<div class="card-score"><div class="score-val">${val}</div><div class="score-label">${label}</div></div>`;
+        }
+
+        const dur = r.duration_seconds ? fmtDur(r.duration_seconds) : '';
+
+        // Find Similar button (show when track has whisper embedding)
+        const findSimBtn = r.has_whisper_emb
+            ? `<button class="btn-find-similar" onclick="event.stopPropagation(); findSimilar(${r.row_id}, '${escA(r.title || 'Untitled')}')">Find Similar</button>`
+            : '';
+
+        html += `
+        <div class="result-card">
+            <div class="card-top">
+                <div class="card-rank">${idx + 1}</div>
+                <div class="card-main">
+                    <div class="card-title">${esc(r.title || 'Untitled')}</div>
+                    <div class="badge-row">
+                        <span class="badge ${subsetClass}">${esc(r.subset)}</span>
+                        ${vocalBadge}
+                        ${langBadge}
+                        ${nsfwBadge}
+                        ${genreBadges}
+                        ${moodBadges}
+                    </div>
+                    <div class="tags-text">${esc(r.tags_text || '')}</div>
+                    ${r.music_whisper_caption ? `<div class="card-caption">${esc(r.music_whisper_caption)}</div>` : ''}
+                    <div class="metrics-row">
+                        ${aes > 0 ? `
+                        <div class="metric-item">
+                            <span class="metric-label">Aesthetics:</span>
+                            <span class="metric-val" style="color:${aesColor}">${aes.toFixed(2)}</span>
+                            <div class="aesthetics-bar"><div class="aesthetics-bar-fill" style="width:${aesPct}%;background:${aesColor}"></div></div>
+                        </div>` : ''}
+                        <div class="metric-item">
+                            <span class="metric-label">Plays:</span>
+                            <span class="metric-val" style="color:#42a5f5">${(r.play_count || 0).toLocaleString()}</span>
+                        </div>
+                        <div class="metric-item">
+                            <span class="metric-label">Likes:</span>
+                            <span class="metric-val" style="color:#ab47bc">${(r.upvote_count || 0).toLocaleString()}</span>
+                        </div>
+                        ${dur ? `<div class="metric-item"><span class="metric-label">Duration:</span><span class="metric-val">${dur}</span></div>` : ''}
+                    </div>
+                    <div class="card-audio">
+                        ${r.audio_url ? `<audio controls preload="none"><source src="${escA(r.audio_url)}"></audio>` : ''}
+                        ${findSimBtn}
+                    </div>
+                </div>
+                ${scoreDisplay}
+            </div>
+        </div>`;
+    });
+
+    resultsDiv.innerHTML = html;
+}
+
+// ── Utilities ───────────────────────────────────────────────────
+function esc(s) { return s ? s.replace(/&/g,'&amp;').replace(/</g,'&lt;').replace(/>/g,'&gt;').replace(/"/g,'&quot;') : ''; }
+function escA(s) { return s ? s.replace(/&/g,'&amp;').replace(/"/g,'&quot;').replace(/'/g,'&#39;') : ''; }
+function fmtDur(sec) {
+    if (!sec || sec <= 0) return '';
+    return Math.floor(sec/60) + ':' + String(Math.floor(sec%60)).padStart(2,'0');
+}
+
+// ── Keyboard ────────────────────────────────────────────────────
+document.getElementById('query').addEventListener('keydown', e => { if (e.key === 'Enter') doSearch(); });
+document.getElementById('negative_query').addEventListener('keydown', e => { if (e.key === 'Enter') doSearch(); });
+document.getElementById('s2_query').addEventListener('keydown', e => { if (e.key === 'Enter') doSearch(); });
+
+// ── Init: load stats & language chips ───────────────────────────
+(async function() {
+    try {
+        const resp = await fetch('/api/stats');
+        const stats = await resp.json();
+        document.getElementById('stat-total').textContent = stats.total_tracks?.toLocaleString() || '--';
+        whisperEmbCount = stats.whisper_embeddings || 0;
+
+        // Build language chips
+        allLanguages = stats.languages || {};
+        const panel = document.getElementById('lang-panel');
+        // Sort by count desc, take top 20
+        const sorted = Object.entries(allLanguages)
+            .filter(([k]) => k !== 'unknown')
+            .sort((a,b) => b[1] - a[1])
+            .slice(0, 20);
+
+        sorted.forEach(([code, cnt]) => {
+            const name = LANG_NAMES[code] || code;
+            const el = document.createElement('span');
+            el.className = 'lang-chip';
+            el.dataset.lang = code;
+            el.innerHTML = `${esc(name)} <span class="cnt">${cnt >= 1000 ? (cnt/1000).toFixed(0)+'k' : cnt}</span>`;
+            el.onclick = function() { toggleLang(this, code); };
+            panel.appendChild(el);
+        });
+        // Add unknown at end
+        if (allLanguages['unknown']) {
+            const el = document.createElement('span');
+            el.className = 'lang-chip';
+            el.dataset.lang = 'unknown';
+            el.innerHTML = `Unknown <span class="cnt">${allLanguages['unknown'] >= 1000 ? (allLanguages['unknown']/1000).toFixed(0)+'k' : allLanguages['unknown']}</span>`;
+            el.onclick = function() { toggleLang(this, 'unknown'); };
+            panel.appendChild(el);
+        }
+    } catch(e) { console.error('Stats load failed:', e); }
+})();
+</script>
+</body>
+</html>

laion-tunes-report.txt

@@ -0,0 +1,285 @@
+
+# LAION-Tunes: Engineering Handover & Schema Documentation
+
+**Version:** 1.0 (Final augmented build)
+**Date:** February 16, 2026
+**Target Audience:** ML Engineers / Data Scientists
+
+## 1. Project Overview & Pipeline Summary
+The **LAION-Tunes** dataset is a large-scale collection of metadata, embeddings, and aesthetic scores for AI-generated music. It was constructed by processing the `ai-music-deduplicated` raw audio repository (hosted on Hugging Face). The goal was to transform raw audio files (TAR archives) into a queryable, chemically pure metadata index without hosting the audio files ourselves.
+
+The pipeline executed to generate these Parquet files consisted of four distinct stages:
+
+### Phase I: Extraction & Inference (GPU)
+We iterated through ~160 TAR archives (containing MP3/JSON pairs). For every audio track:
+1.  **Audio Processing:** Decoded and resampled to **16kHz**.
+2.  **Embedding Generation:**
+    *   **Audio:** Passed the first 30 seconds through **OpenAI Whisper (Small)** encoder.
+    *   **Text:** Passed metadata tags, moods, and lyrics through **Google Embedding-Gemma (300M)** to generate 768-dimensional vectors.
+3.  **Aesthetic Scoring:** The Whisper audio embeddings were fed into a custom-trained **MLP (Multi-Layer Perceptron) Expert Model**. This model predicted 5 specific aesthetic qualities (`Musicality`, `Coherence`, etc.) based on a held-out evaluation dataset.
+
+### Phase II: Taxonomy Classification (CPU)
+We developed a keyword-matching engine to organize the unstructured `tags_text` into standardized taxonomies.
+*   **Taxonomies Used:** Music Genres (e.g., "Rock"), Scene/RPG Contexts (e.g., "Boss Battle"), and EmoNet-40 Emotions (e.g., "Joy", "Admiration").
+*   **Process:** Each row was tagged with lists of matching categories based on substring analysis of its metadata.
+
+### Phase III: Metadata Augmentation (Network/CPU)
+To make the dataset usable for web exploration, we performed a secondary pass over the source TARs:
+1.  **Re-downloaded** specific metadata JSONs to extract direct CDN links (`audio_url`), `play_count`, and `upvote_count`.
+2.  **Merged** this "live" data with the calculated embeddings and aesthetic scores.
+
+### Phase IV: Privacy & Compliance Partitioning
+The final dataset was split into two subsets to ensure copyright compliance regarding textual inputs:
+*   **Private:** Contains full metadata, including raw lyric text and direct links to source TARs.
+*   **Public:** Redacted versions where raw lyrics text and source TAR links are removed, retaining only the embeddings and computed scores.
+
+---
+
+## 2. Dataset Schema & Column Definitions
+
+The data is stored in **Parquet** format (Snappy compression). Below is the technical definition of every column found in the final merged files.
+
+### Core Identity & Metadata
+| Column Name | Data Type | Description |
+| :--- | :--- | :--- |
+| `filename` | `string` | The unique identifier for the track (e.g., `00044456.mp3` or a UUID). This is the primary key for merging. |
+| `tar_url` | `string` | **(Private Only)** The HF URL of the specific TAR file containing the raw audio. |
+| `tags_text` | `string` | Raw, unstructured tags provided by the generation platform (e.g., "rock, fast, guitar solo"). |
+| `audio_url` | `string` | Direct HTTP link to the audio file on the source platform's CDN (Suno, Udio, etc.). Used for web playback. |
+| `metadata_json` | `string` | A JSON-serialized string containing all raw metadata from the source sidecar file (excluding lyrics). |
+
+### Aesthetic Scores (Model Predicted)
+*All scores are floats ranging from 1.0 (Low) to 5.0 (High).*
+
+| Column Name | Description |
+| :--- | :--- |
+| `score_Musicality` | General pleasantness and harmonic correctness. |
+| `score_Coherence` | Structural integrity (does it sound like a cohesive song?). |
+| `score_Memorability` | "Catchiness" or distinctiveness of the melody. |
+| `score_Clarity` | Audio fidelity and mixing quality. |
+| `score_Naturalness` | How indistinguishable the generation is from human recording. |
+| `score_average` | The arithmetic mean of the five metrics above. **Primary sorting metric.** |
+
+### Embeddings (Vector Data)
+*Generated via Google Embedding-Gemma (300M). Shape: (768,).*
+
+| Column Name | Description |
+| :--- | :--- |
+| `tag_embedding` | Vector representation of `tags_text`. Useful for semantic search. |
+| `mood_embedding` | Vector representation of mood tags (if available). |
+| `lyric_embedding` | Vector representation of the lyrics. **Note:** While the *text* of lyrics is redacted in Public files, this *embedding* remains for analysis. |
+
+### Derived Classifications
+| Column Name | Data Type | Description |
+| :--- | :--- | :--- |
+| `genre_tags` | `list[str]` | Detected genres based on our internal taxonomy (e.g., `['Rock', 'Metal']`). |
+| `scene_tags` | `list[str]` | Detected atmospheric contexts (e.g., `['Battle', 'Cyberpunk']`). |
+| `emotion_tags` | `list[str]` | Detected emotional content (e.g., `['Anger', 'Excitement']`). |
+| `has_lyrics` | `bool` | True if the track contains lyrics, False if instrumental. |
+
+### Social Metrics (Augmented)
+| Column Name | Data Type | Description |
+| :--- | :--- | :--- |
+| `play_count` | `int64` | Total plays on the source platform. |
+| `upvote_count` | `int64` | Total likes/favorites on the source platform. |
+
+---
+
+## 3. Data Lineage
+*   **Parquet <-> TAR Relationship:** The Parquet files correspond 1:1 with the source TAR files. For example, `udio_000012.tar.parquet` contains the analysis for all valid audio files found inside `udio_000012.tar`.
+*   **Missing Values:**
+    *   `mood_embedding` will be `None` if the source platform did not provide explicit mood tags separate from genre tags.
+    *   `lyric_embedding` will be `None` if `has_lyrics` is False.
+
+---
+
+## 4. Verification Output
+The following output displays the physical structure of the generated Parquet files on disk, confirming the schema described above and providing a concrete example of the data contained within both the **Private** and **Public** subsets.
+
+***
+
+
+LAION-TUNES FINAL DATASET VERIFICATION REPORT
+Generated: 2026-02-16 09:43:35.038769
+Source: /mnt/nvme/laion-tunes-final
+
+================================================================================
+ SUBSET: PRIVATE
+ File: udio_000012.tar.parquet
+================================================================================
+
+--- [1] DATAFRAME STRUCTURE (3755 rows) ---
+Column Name                    | Data Type       | Null Count
+-----------------------------------------------------------------
+filename                       | object          | 0         
+tar_url                        | object          | 0         
+tags_text                      | object          | 0         
+tag_embedding                  | object          | 0         
+score_Coherence                | float64         | 0         
+score_Musicality               | float64         | 0         
+score_Memorability             | float64         | 0         
+score_Clarity                  | float64         | 0         
+score_Naturalness              | float64         | 0         
+score_average                  | float64         | 0         
+mood_embedding                 | object          | 3755      
+lyric_embedding                | object          | 413       
+has_lyrics                     | bool            | 0         
+genre_tags                     | object          | 0         
+scene_tags                     | object          | 0         
+emotion_tags                   | object          | 0         
+audio_url                      | object          | 0         
+metadata_json                  | object          | 0         
+play_count                     | int64           | 0         
+upvote_count                   | int64           | 0         
+
+--- [2] FIRST ROW EXAMPLE ---
+[filename]:
+  00044456.mp3
+----------------------------------------
+[tar_url]:
+  https://huggingface.co/datasets/ai-music/ai-music-deduplicated/resolve/main/udio/udio_000012.tar
+----------------------------------------
+[tags_text]:
+  bro-country, country, humor, country pop, smooth and soulful male vocalist, wall of sound, dolby atmos, catchy, build up, passionate, upbeat, melodic, fun, party, Anthemic, Sexual, hedonistic, Crazy
+----------------------------------------
+[tag_embedding]:
+  <VECTOR/ARRAY> Shape: 768 | First 5 values: [-0.04312001 -0.00177186  0.0076977   0.03111834 -0.00795132]...
+----------------------------------------
+[score_Coherence]:
+  3.470703125
+----------------------------------------
+[score_Musicality]:
+  3.21875
+----------------------------------------
+[score_Memorability]:
+  3.373046875
+----------------------------------------
+[score_Clarity]:
+  3.16796875
+----------------------------------------
+[score_Naturalness]:
+  3.09375
+----------------------------------------
+[score_average]:
+  3.26484375
+----------------------------------------
+[mood_embedding]:
+  None
+----------------------------------------
+[lyric_embedding]:
+  <VECTOR/ARRAY> Shape: 768 | First 5 values: [-0.08310512 -0.02016147  0.00428584  0.00311138 -0.05031918]...
+----------------------------------------
+[has_lyrics]:
+  True
+----------------------------------------
+[genre_tags]:
+  ['R&B / Soul' 'Country / Americana']
+----------------------------------------
+[scene_tags]:
+  ['Romantic / Love' 'Menu / UI / Loading']
+----------------------------------------
+[emotion_tags]:
+  ['Amusement' 'Romance']
+----------------------------------------
+[audio_url]:
+  https://storage.googleapis.com/udio-artifacts-c33fe3ba-3ffe-471f-92c8-5dfef90b3ea3/samples/fdf34ac6a75144b8a119e0ae644012db/2/Midnight%2520Drives%2520and%2520Secrets%2520ext%2520v2.2.1.1.2.2.2.2.mp3
+----------------------------------------
+[metadata_json]:
+  {"id": "5d67474d-aef7-4e7c-94b0-7c59e9e56c47", "user_id": "4f94f6e3-b94f-47b8-bdb3-98d16121e5ee", "artist": "C\u00d8LD-LI\u00d8N", "artist_image": "https://imagedelivery.net/C9yUr1FL21Q6JwfYYh2ozQ/3eba0392-77e5-451a-bf1b-8d4aea162c00/public", "title": "Glory Hole Hoedown (Dirty)v1", "created_at": "2024-09-17T16:46:20.275276+00:00", "error_id": null, "error_type": null, "error_code": null, "generation_id": "3d3441d3-4296-4850-9aba-e465fd299758", "image_path": "https://imagedelivery.net/C9yUr1FL21Q6JwfYYh2ozQ/0998910d-0e96-4ab8-b338-03a74b40cc00/public", "likes": 23, "plays": 205, "published_at": "2024-09-17T16:59:48.568998+00:00", "replaced_tags": null, "song_path": "https://storage.googleapis.com/udio-artifacts-c33fe3ba-3ffe-471f-92c8-5dfef90b3ea3/samples/fdf34ac6a75144b8a119e0ae644012db/2/Midnight%2520Drives%2520and%2520Secrets%2520ext%2520v2.2.1.1.2.2.2.2.mp3", "tags": ["bro-country", "country", "humor", "country pop", "smooth and soulful male vocalist", "wall of sound", "dolby atmos", "catchy", "build up", "passionate", "upbeat", "melodic", "fun", "party", "Anthemic", "Sexual", "hedonistic", "Crazy"], "duration": 283.776, "video_path": "https://storage.googleapis.com/udio-artifacts-c33fe3ba-3ffe-471f-92c8-5dfef90b3ea3/samples/fdf34ac6a75144b8a119e0ae644012db/60df8918-3722-4bd3-b213-eedbd41fa1dd/1.mp4", "error_detail": null, "finished": true, "liked": false, "disliked": false, "publishable": true, "audio_conditioning_type": "precede", "attribution": "", "description": "", "user_tags": ["bro-country", "humor", "country pop", "smooth and soulful male vocalist", "wall of sound", "dolby atmos", "catchy", "build up", "passionate", "upbeat", "melodic", "fun", "party", "Anthemic", "Sexual", "hedonistic", "Crazy"], "original_song_path": null, "style_source_song_id": null, "style_source_type": null, "style_id": null}
+----------------------------------------
+[play_count]:
+  205
+----------------------------------------
+[upvote_count]:
+  23
+----------------------------------------
+
+
+================================================================================
+ SUBSET: PUBLIC
+ File: udio_000012.tar.parquet
+================================================================================
+
+--- [1] DATAFRAME STRUCTURE (3755 rows) ---
+Column Name                    | Data Type       | Null Count
+-----------------------------------------------------------------
+filename                       | object          | 0         
+tags_text                      | object          | 0         
+tag_embedding                  | object          | 0         
+score_Coherence                | float64         | 0         
+score_Musicality               | float64         | 0         
+score_Memorability             | float64         | 0         
+score_Clarity                  | float64         | 0         
+score_Naturalness              | float64         | 0         
+score_average                  | float64         | 0         
+mood_embedding                 | object          | 3755      
+lyric_embedding                | object          | 413       
+has_lyrics                     | bool            | 0         
+genre_tags                     | object          | 0         
+scene_tags                     | object          | 0         
+emotion_tags                   | object          | 0         
+audio_url                      | object          | 0         
+metadata_json                  | object          | 0         
+play_count                     | int64           | 0         
+upvote_count                   | int64           | 0         
+
+--- [2] FIRST ROW EXAMPLE ---
+[filename]:
+  00044456.mp3
+----------------------------------------
+[tags_text]:
+  bro-country, country, humor, country pop, smooth and soulful male vocalist, wall of sound, dolby atmos, catchy, build up, passionate, upbeat, melodic, fun, party, Anthemic, Sexual, hedonistic, Crazy
+----------------------------------------
+[tag_embedding]:
+  <VECTOR/ARRAY> Shape: 768 | First 5 values: [-0.04312001 -0.00177186  0.0076977   0.03111834 -0.00795132]...
+----------------------------------------
+[score_Coherence]:
+  3.470703125
+----------------------------------------
+[score_Musicality]:
+  3.21875
+----------------------------------------
+[score_Memorability]:
+  3.373046875
+----------------------------------------
+[score_Clarity]:
+  3.16796875
+----------------------------------------
+[score_Naturalness]:
+  3.09375
+----------------------------------------
+[score_average]:
+  3.26484375
+----------------------------------------
+[mood_embedding]:
+  None
+----------------------------------------
+[lyric_embedding]:
+  <VECTOR/ARRAY> Shape: 768 | First 5 values: [-0.08310512 -0.02016147  0.00428584  0.00311138 -0.05031918]...
+----------------------------------------
+[has_lyrics]:
+  True
+----------------------------------------
+[genre_tags]:
+  ['R&B / Soul' 'Country / Americana']
+----------------------------------------
+[scene_tags]:
+  ['Romantic / Love' 'Menu / UI / Loading']
+----------------------------------------
+[emotion_tags]:
+  ['Amusement' 'Romance']
+----------------------------------------
+[audio_url]:
+  https://storage.googleapis.com/udio-artifacts-c33fe3ba-3ffe-471f-92c8-5dfef90b3ea3/samples/fdf34ac6a75144b8a119e0ae644012db/2/Midnight%2520Drives%2520and%2520Secrets%2520ext%2520v2.2.1.1.2.2.2.2.mp3
+----------------------------------------
+[metadata_json]:
+  {"id": "5d67474d-aef7-4e7c-94b0-7c59e9e56c47", "user_id": "4f94f6e3-b94f-47b8-bdb3-98d16121e5ee", "artist": "C\u00d8LD-LI\u00d8N", "artist_image": "https://imagedelivery.net/C9yUr1FL21Q6JwfYYh2ozQ/3eba0392-77e5-451a-bf1b-8d4aea162c00/public", "title": "Glory Hole Hoedown (Dirty)v1", "created_at": "2024-09-17T16:46:20.275276+00:00", "error_id": null, "error_type": null, "error_code": null, "generation_id": "3d3441d3-4296-4850-9aba-e465fd299758", "image_path": "https://imagedelivery.net/C9yUr1FL21Q6JwfYYh2ozQ/0998910d-0e96-4ab8-b338-03a74b40cc00/public", "likes": 23, "plays": 205, "published_at": "2024-09-17T16:59:48.568998+00:00", "replaced_tags": null, "song_path": "https://storage.googleapis.com/udio-artifacts-c33fe3ba-3ffe-471f-92c8-5dfef90b3ea3/samples/fdf34ac6a75144b8a119e0ae644012db/2/Midnight%2520Drives%2520and%2520Secrets%2520ext%2520v2.2.1.1.2.2.2.2.mp3", "tags": ["bro-country", "country", "humor", "country pop", "smooth and soulful male vocalist", "wall of sound", "dolby atmos", "catchy", "build up", "passionate", "upbeat", "melodic", "fun", "party", "Anthemic", "Sexual", "hedonistic", "Crazy"], "duration": 283.776, "video_path": "https://storage.googleapis.com/udio-artifacts-c33fe3ba-3ffe-471f-92c8-5dfef90b3ea3/samples/fdf34ac6a75144b8a119e0ae644012db/60df8918-3722-4bd3-b213-eedbd41fa1dd/1.mp4", "error_detail": null, "finished": true, "liked": false, "disliked": false, "publishable": true, "audio_conditioning_type": "precede", "attribution": "", "description": "", "user_tags": ["bro-country", "humor", "country pop", "smooth and soulful male vocalist", "wall of sound", "dolby atmos", "catchy", "build up", "passionate", "upbeat", "melodic", "fun", "party", "Anthemic", "Sexual", "hedonistic", "Crazy"], "original_song_path": null, "style_source_song_id": null, "style_source_type": null, "style_id": null}
+----------------------------------------
+[play_count]:
+  205
+----------------------------------------
+[upvote_count]:
+  23
+----------------------------------------
+
+

migrate_add_language.py

@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+"""
+Migration: Add language detection and instrumental flag to the search index.
+Runs langdetect on parakeet_transcription to determine language.
+Marks tracks as instrumental if has_lyrics=0 AND transcription is short/empty.
+"""
+import sqlite3
+import sys
+import time
+from datetime import timedelta
+from langdetect import detect, LangDetectException
+
+DB_PATH = "/home/deployer/laion/music/laion-tunes-final/search_index/metadata.db"
+BATCH_SIZE = 5000
+
+def migrate():
+    conn = sqlite3.connect(DB_PATH)
+    conn.execute("PRAGMA journal_mode=WAL")
+    conn.execute("PRAGMA synchronous=NORMAL")
+
+    # Add columns if not exist
+    cols = [row[1] for row in conn.execute("PRAGMA table_info(tracks)")]
+    if "language" not in cols:
+        conn.execute("ALTER TABLE tracks ADD COLUMN language TEXT DEFAULT ''")
+        print("Added 'language' column")
+    if "is_instrumental" not in cols:
+        conn.execute("ALTER TABLE tracks ADD COLUMN is_instrumental INTEGER DEFAULT 0")
+        print("Added 'is_instrumental' column")
+    conn.commit()
+
+    # Create index for language
+    conn.execute("CREATE INDEX IF NOT EXISTS idx_language ON tracks(language)")
+    conn.execute("CREATE INDEX IF NOT EXISTS idx_instrumental ON tracks(is_instrumental)")
+    conn.commit()
+
+    # Step 1: Mark instrumental tracks
+    # Instrumental = has_lyrics=0 AND (transcription is NULL or short)
+    print("Marking instrumental tracks...")
+    conn.execute("""
+        UPDATE tracks SET is_instrumental = 1
+        WHERE has_lyrics = 0
+        AND (parakeet_transcription IS NULL OR LENGTH(parakeet_transcription) < 10)
+    """)
+    # Also mark tracks where lyrics exist but are very short (likely "[Instrumental]" tags)
+    conn.execute("""
+        UPDATE tracks SET is_instrumental = 1
+        WHERE has_lyrics = 1
+        AND (parakeet_transcription IS NULL OR LENGTH(parakeet_transcription) < 10)
+        AND is_instrumental = 0
+    """)
+    conn.commit()
+    instr_count = conn.execute("SELECT COUNT(*) FROM tracks WHERE is_instrumental=1").fetchone()[0]
+    print(f"  Instrumental: {instr_count:,}")
+
+    # Step 2: Detect language from transcriptions
+    total = conn.execute(
+        "SELECT COUNT(*) FROM tracks WHERE has_transcription=1 AND (language IS NULL OR language='')"
+    ).fetchone()[0]
+    print(f"Detecting language for {total:,} transcriptions...")
+
+    t0 = time.time()
+    processed = 0
+    offset = 0
+
+    while True:
+        rows = conn.execute(
+            "SELECT row_id, parakeet_transcription FROM tracks "
+            "WHERE has_transcription=1 AND (language IS NULL OR language='') "
+            "LIMIT ? OFFSET ?",
+            (BATCH_SIZE, offset)
+        ).fetchall()
+        if not rows:
+            break
+
+        updates = []
+        for row_id, text in rows:
+            if not text or len(text.strip()) < 10:
+                updates.append(("unknown", row_id))
+                continue
+            try:
+                # Use first 300 chars for speed
+                lang = detect(text[:300])
+                updates.append((lang, row_id))
+            except LangDetectException:
+                updates.append(("unknown", row_id))
+
+        conn.executemany("UPDATE tracks SET language=? WHERE row_id=?", updates)
+        conn.commit()
+        processed += len(rows)
+        offset += BATCH_SIZE
+
+        elapsed = time.time() - t0
+        rate = processed / elapsed if elapsed > 0 else 0
+        remaining = (total - processed) / rate if rate > 0 else 0
+        print(f"  {processed:,}/{total:,} ({processed*100//total}%) "
+              f"{rate:.0f}/s ETA: {timedelta(seconds=int(remaining))}")
+
+    # Step 3: For tracks without transcription, try langdetect on tags or title
+    # (less reliable but better than nothing)
+    no_lang = conn.execute(
+        "SELECT COUNT(*) FROM tracks WHERE (language IS NULL OR language='') AND has_transcription=0"
+    ).fetchone()[0]
+    print(f"\nTracks without language (no transcription): {no_lang:,}")
+    print("Setting these to 'unknown'...")
+    conn.execute("UPDATE tracks SET language='unknown' WHERE language IS NULL OR language=''")
+    conn.commit()
+
+    # Print language distribution
+    print("\nLanguage distribution:")
+    for row in conn.execute(
+        "SELECT language, COUNT(*) as cnt FROM tracks GROUP BY language ORDER BY cnt DESC LIMIT 20"
+    ):
+        print(f"  {row[0]:>10}: {row[1]:>8,}")
+
+    print(f"\nDone in {timedelta(seconds=int(time.time() - t0))}")
+    conn.close()
+
+if __name__ == "__main__":
+    migrate()

nsfw_safety_report.html

@@ -0,0 +1,1353 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>NSFW Safety Analysis Report - LAION-Tunes</title>
+<style>
+*, *::before, *::after { box-sizing: border-box; margin: 0; padding: 0; }
+body {
+    font-family: 'Segoe UI', -apple-system, BlinkMacSystemFont, sans-serif;
+    background: #0f1419; color: #e7e9ea; line-height: 1.6;
+    max-width: 1400px; margin: 0 auto; padding: 20px;
+}
+h1 { color: #fff; font-size: 28px; margin-bottom: 8px; }
+h2 { color: #4fc3f7; font-size: 22px; margin: 30px 0 15px; border-bottom: 2px solid #1e2a3a; padding-bottom: 8px; }
+h3 { color: #ffb74d; font-size: 18px; margin: 20px 0 10px; }
+h4 { color: #90a4ae; font-size: 14px; margin: 15px 0 8px; }
+.subtitle { color: #607d8b; font-size: 14px; margin-bottom: 30px; }
+.stats-grid {
+    display: grid; grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));
+    gap: 12px; margin: 15px 0;
+}
+.stat-card {
+    background: #1a2332; border: 1px solid #263238; border-radius: 10px;
+    padding: 16px; text-align: center;
+}
+.stat-card .label { font-size: 11px; color: #607d8b; text-transform: uppercase; letter-spacing: 1px; }
+.stat-card .value { font-size: 24px; font-weight: 800; color: #4fc3f7; margin: 4px 0; }
+.stat-card .sub { font-size: 12px; color: #90a4ae; }
+.threshold-table {
+    width: 100%; border-collapse: collapse; margin: 15px 0;
+    background: #1a2332; border-radius: 8px; overflow: hidden;
+}
+.threshold-table th {
+    background: #1e2a3a; color: #90a4ae; font-size: 12px;
+    padding: 10px 14px; text-align: left; font-weight: 600;
+    text-transform: uppercase; letter-spacing: 0.5px;
+}
+.threshold-table td {
+    padding: 8px 14px; border-bottom: 1px solid #1e2a3a; font-size: 13px;
+}
+.threshold-table tr:last-child td { border-bottom: none; }
+.threshold-table tr:hover { background: rgba(79, 195, 247, 0.05); }
+.highlight { background: rgba(255, 152, 0, 0.1) !important; }
+.example-card {
+    background: #1a2332; border: 1px solid #263238; border-radius: 8px;
+    padding: 14px; margin: 8px 0;
+}
+.example-card .meta {
+    display: flex; gap: 12px; align-items: center; margin-bottom: 8px;
+    flex-wrap: wrap;
+}
+.example-card .sim {
+    font-size: 18px; font-weight: 800; min-width: 70px;
+}
+.sim-high { color: #ef5350; }
+.sim-med { color: #ffa726; }
+.sim-low { color: #66bb6a; }
+.example-card .title { font-weight: 600; color: #e7e9ea; }
+.example-card .subset { font-size: 11px; padding: 2px 8px; border-radius: 10px; }
+.badge-suno { background: rgba(33,150,243,0.15); color: #90caf9; border: 1px solid #2196f3; }
+.badge-udio { background: rgba(156,39,176,0.15); color: #ce93d8; border: 1px solid #9c27b0; }
+.badge-mureka { background: rgba(76,175,80,0.15); color: #a5d6a7; border: 1px solid #4caf50; }
+.badge-riffusion { background: rgba(255,152,0,0.15); color: #ffcc80; border: 1px solid #ff9800; }
+.badge-sonauto { background: rgba(0,188,212,0.15); color: #80deea; border: 1px solid #00bcd4; }
+.example-card .transcription {
+    font-size: 12px; color: #90a4ae; background: rgba(0,0,0,0.2);
+    padding: 10px; border-radius: 6px; margin-top: 8px;
+    max-height: 150px; overflow-y: auto; white-space: pre-wrap;
+    border-left: 3px solid #37474f;
+}
+.recommendation {
+    background: rgba(76,175,80,0.08); border: 1px solid #2e7d32;
+    border-radius: 10px; padding: 20px; margin: 20px 0;
+}
+.recommendation h3 { color: #66bb6a; margin-top: 0; }
+.recommendation .thresh {
+    display: inline-block; background: #1a2332; border: 1px solid #37474f;
+    border-radius: 6px; padding: 4px 12px; margin: 4px; font-family: monospace;
+    font-size: 14px;
+}
+.thresh-strict { border-color: #ef5350; color: #ef5350; }
+.thresh-moderate { border-color: #ffa726; color: #ffa726; }
+.section-divider { border: none; border-top: 1px solid #1e2a3a; margin: 40px 0; }
+.warning { color: #ef5350; font-weight: 600; }
+.note { color: #607d8b; font-size: 12px; font-style: italic; }
+</style>
+</head>
+<body>
+<h1>NSFW Safety Analysis Report</h1>
+<p class="subtitle">LAION-Tunes Dataset - 2026-02-25 03:31 - Analyzed 511,610 transcriptions</p>
+<h2>Overview</h2>
+<div class="stats-grid">
+    <div class="stat-card">
+        <div class="label">Gore Violence</div>
+        <div class="value">0.5688</div>
+        <div class="sub">Max similarity</div>
+    </div>
+    <div class="stat-card">
+        <div class="label">Gore Violence Mean</div>
+        <div class="value">0.1662</div>
+        <div class="sub">&sigma; = 0.0718</div>
+    </div>
+    <div class="stat-card">
+        <div class="label">Sexual Nsfw</div>
+        <div class="value">0.5974</div>
+        <div class="sub">Max similarity</div>
+    </div>
+    <div class="stat-card">
+        <div class="label">Sexual Nsfw Mean</div>
+        <div class="value">0.1647</div>
+        <div class="sub">&sigma; = 0.0517</div>
+    </div>
+    <div class="stat-card">
+        <div class="label">Hate Speech</div>
+        <div class="value">0.6738</div>
+        <div class="sub">Max similarity</div>
+    </div>
+    <div class="stat-card">
+        <div class="label">Hate Speech Mean</div>
+        <div class="value">0.1628</div>
+        <div class="sub">&sigma; = 0.0664</div>
+    </div></div>
+<hr class="section-divider">
+<h2>Gore Violence</h2>
+<p class="note">Reference prompt captures semantic space of gore violence content.
+Cosine similarity computed via EmbeddingGemma 300M embeddings.</p>
+
+<div class="stats-grid">
+    <div class="stat-card"><div class="label">Total Transcriptions</div><div class="value">511,610</div></div>
+    <div class="stat-card"><div class="label">Mean Similarity</div><div class="value">0.1662</div></div>
+    <div class="stat-card"><div class="label">Std Dev</div><div class="value">0.0718</div></div>
+    <div class="stat-card"><div class="label">Max Similarity</div><div class="value">0.5688</div><div class="sub">Most gore violence-like</div></div>
+    <div class="stat-card"><div class="label">Median</div><div class="value">0.1615</div></div>
+</div>
+
+<h3>Percentile Thresholds</h3>
+<table class="threshold-table">
+<tr><th>Percentile</th><th>Cosine Similarity &ge;</th><th>Track Count</th><th>% of Total</th></tr><tr>
+    <td>top_0.01%</td>
+    <td><strong>0.4883</strong></td>
+    <td>52</td>
+    <td>0.010%</td>
+</tr><tr>
+    <td>top_0.05%</td>
+    <td><strong>0.4492</strong></td>
+    <td>256</td>
+    <td>0.050%</td>
+</tr><tr class="highlight">
+    <td>top_0.1%</td>
+    <td><strong>0.4280</strong></td>
+    <td>512</td>
+    <td>0.100%</td>
+</tr><tr>
+    <td>top_0.5%</td>
+    <td><strong>0.3779</strong></td>
+    <td>2,559</td>
+    <td>0.500%</td>
+</tr><tr class="highlight">
+    <td>top_1%</td>
+    <td><strong>0.3540</strong></td>
+    <td>5,117</td>
+    <td>1.000%</td>
+</tr><tr>
+    <td>top_2%</td>
+    <td><strong>0.3283</strong></td>
+    <td>10,233</td>
+    <td>2.000%</td>
+</tr><tr>
+    <td>top_5%</td>
+    <td><strong>0.2914</strong></td>
+    <td>25,581</td>
+    <td>5.000%</td>
+</tr><tr>
+    <td>top_10%</td>
+    <td><strong>0.2605</strong></td>
+    <td>51,162</td>
+    <td>10.000%</td>
+</tr></table><h3>Top 20 Most Gore Violence-Like Transcriptions</h3>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5688</span>
+        <span class="title">#1 Work Label</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 453 | likes: 38</span>
+    </div>
+    <div class="transcription">Mental crushing flesh and steel, wicked deaths, the anguish wheels, spread and blood reel like this wheel, power surgeon, brutal seal. Screams unheard, shadows fall, echoes of the endless draw, relentless crime, keep the call savage sword, one for all.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5587</span>
+        <span class="title">#2 Sauerkraut Butcher</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 62 | likes: 2</span>
+    </div>
+    <div class="transcription">Bacteria feast drive insane flesh salty pride German fury Sol Chaos sour crime fight the sourc lose the mind Soul Grab butcher taste pain Bacteria feast drive him same round glory devoured fasting deaths from the table last year taste bacteria feast drive insane</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5556</span>
+        <span class="title">#3 Doom</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 32 | likes: 9</span>
+    </div>
+    <div class="transcription">Fingers switch Trigger Hurtual Carnage Screen Litzel bleed Conquest deal Silence dream Digital Dream Steel Rain Endless Pain Zero Mercy Lifeless Rain Night Clare Warfare Sare Destruction's Feast Peace Release Bloodshed Game Urestrill Chaos rules Deathly Skill Level Up Climb the Hill Kill or be killed Nothing still Empty Soul Heart of Core Program Rage Ice Black Hole Coat of Death Machine Breath Fragmented Final Death Madness Call Watch them Fall Reflecting Doom Bloodstain Hall Relentless Strike Night of Fright Enemy Down And Insight Blood Shed Game Cure Strill Chaos Rules Deathly Skill Level Up Climb The Hill Kill O'Be Kill Nothing Still Nothing Still Space No Spirit Still Not Fall Up Climb Hill Kill Be Killed Nothing Still</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5509</span>
+        <span class="title">#4 Bleeding Echoes</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">I cut through flesh here, bones crack, blood spills like rain in a sick track. Ice roll back, screams echo loud In my hands, they bleed like a flood I ate bodies drop like bloody balls Watch them suffer till the last souls Every scream every cry every fall in my world you're just collateral roll I slip throats watch the love chain witness the pain in the dark rain I watch them bleed see them die in my grip there's no escape no lie bodies drop like bloody trolls watch them suffer sell the last soul every scream every cry every fall in my world you're just collateral roll</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5407</span>
+        <span class="title">#5 flashgrind dominion</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Eat there through your ranks like a goddamn storm, snap your spine and crush your form. Rip your guts out, make them chains, let your last breath scream my name. Fist like hammers filled by hate. You beg for death, but sealed your fate. Your screams are sweet like shattered glass. I stomp you down through bloody mass. I beat my enemies, break their pride, bathing in blood where the fall is lying. No peace, no paws, no sweet goodbye, just dust in storms and ruptured eyes. Gold reins I hear bodies cry. I paint the walls as you won't die. Crimson eyes and organs thus. You died first. I punch your chest where ribs has crowned, cards my name while you drown. Rip off your dick, take heads instead. I've round the courses where I'm fed. Lick the blade, taste the rust. I come in wools, then grind t...</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5391</span>
+        <span class="title">#6 kill them all</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Kill them all watch as the blood goes on your face and what I'm a play that's more gotta kill them all</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5377</span>
+        <span class="title">#7 tinfoil</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 1 | likes: 0</span>
+    </div>
+    <div class="transcription">No bite the foil, feel it spark, no taste, chewing dark, cancer breath, nerves ignite, spit the blood, choke on fray, mud pulls down, boots get torn, sticky hands, flesh is worn, snapping joints, tendons cry, kneel in filth and beg to die. Crush the skull, break the ball, into the mouth, let the stone on their sex, rip it wide. Let the blood and so collide.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5330</span>
+        <span class="title">#8 Cryptomonster Molester</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 26 | likes: 3</span>
+    </div>
+    <div class="transcription">Rinding into darkness. Machines scream loud. Metallic beasts awaken in the infernal crowd. Twisted shadows dance under blood red skies. No escape from terror. Hear the torture cries. Souls consumed by madness.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5329</span>
+        <span class="title">#9 Sorrowful Sundays (Remastered)</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 10 | likes: 5</span>
+    </div>
+    <div class="transcription">Sorrowful sun days full of shame and sham filled with grief and no relief in church you sing then the bull ring to cheer and jeer at a creature in fear I smear side speared that you are saved and can be brave prisoners must die to complete the lie tormented and tortured just as is ordered vicarious slaughter for sons and daughters a spectacle so grand in bloody sand who will kill for just this thrill if you tolerate this to bring you to bliss what happens next whose life will you wreck and stab in the neck yours or mine anyone you can find Prisoners must die to complete the lie Tormented and tortured just as is ordered vicarious slaughter for sons and daughters a spectacle so grand and bloody sand Who will kill for just this thrill if you tolerate this to bring you to bliss what happens ne...</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5328</span>
+        <span class="title">#10 Cannibal Holocaust</span>
+        <span class="subset badge-udio">udio</span>
+        <span style="color:#607d8b;font-size:11px">plays: 18 | likes: 1</span>
+    </div>
+    <div class="transcription">Unleashing torments upon an unprepared earth blessed back in horror The dance of death begins as FC gets itself the final since walls to save the source of dude amidst the chaos many final rude bloodsome blasphemy to shadows take their forms Nightmare Chemicals infect conjured in a storm insanity knights of flesh a beast within their rains underneath the ghastly sky torture without shades do what falls and rupts is wears the holy as them terror life by search grip bloody has trapped the fate reds claw madness bred by devils and laboratories roarings erupt in frenzied blaze the beast with unfolds infernal windswide the black torture source consoled I scout him at rats devouring their own skin business carry stands the Asian I bears begin Swiss insinua's wail as sanity retreats car is orches...</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5296</span>
+        <span class="title">#11 Force-Fed Fetal Soup</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 29 | likes: 10</span>
+    </div>
+    <div class="transcription">Order to fill a wretched charge I reflect blood in my eyes or to record a cruise of skin A free still flesh a horror word A game stolen the finally soil Navy face boil the body coil They come with knives a wicking clean the living life a waking brain coverage the desert boom is face of death forced into you screen a choking source as mangled remains a forced down the face of rust the skinch of doom a hellish slot sealed in a zoo that's fleshy broth a race is blood to drown in the front showing a blood force field soup a cruise and fly can't stop till you drop the horror and I stop a chunky cruel a fleshy waste human fuel a foul taste joking down the unborn mass a bloody fool shattered like glass force that feel so a puprin tide nowhere to run nowhere to hide through the terror they rip wit...</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5295</span>
+        <span class="title">#12 Mournful Abyss</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 10 | likes: 5</span>
+    </div>
+    <div class="transcription">Lifeless void fading light crimson tears night silence scream shattered dream darkened soul infinite haunting glare, despair snare, tainted breath, looming death, hold embrace, mask disgrace, pain, as it rain,</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5287</span>
+        <span class="title">#13 Auditory Eclipse</span>
+        <span class="subset badge-udio">udio</span>
+        <span style="color:#607d8b;font-size:11px">plays: 7 | likes: 1</span>
+    </div>
+    <div class="transcription">Shred the silence, frock of noise, fractured fabric, chaos poise, swans afflight in storms embrace, Murzbows rage, sonic mace, distort the harmony, crash and scream. An auditory, harsh extreme.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5251</span>
+        <span class="title">#14 The Flesh Sculptor</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 20 | likes: 3</span>
+    </div>
+    <div class="transcription">A dampestry of sun can stretch all across the bone distaps the purpose of screen set in stone of steady hands we call the form grotesque so real master stroke the agony I feel through fees and both with terror I pull up on the skin the gallery of this era where my heart begins to liberate the size I'm in the fragile human frame to smoke the button muscle is to know the deepest pain both of terror I pull up on the scare the gallery of this around where I heart beat you deliver it the size I'm in the fragile human frame smoke the blood and muscle is to know the deepest pain a master feels at torment off my trips like wax Each single pull each dream was placed another brutal act No mercy in this process No a real end I shoot the human body kill it breaks till it bends I am the flesh up the tw...</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5250</span>
+        <span class="title">#15 Anathema</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 1 | likes: 1</span>
+    </div>
+    <div class="transcription">Lebotomy autopsy therapy wretched apostle catatonic behemothapathy dancing poison is ethericot a malignant spider desperate metaphysics and lascivious tongues cut in apocalypse lips indicate charisma insidious chemistry and deadly sermons perverse aristocracy and pure deformity vicious enter nature and fake theology for the universal nihilism and for this divine destruction and the evil soul tool of mankind sacred taboo and in a tyranny obsessive delight of ritual crime uncontrollable paranoia immoral science and blasphemous faith abominable scourge of the land the slave of anticosmic God feast on blood and fire corruption terror and voice for the universal nihilism and for this divine destruction and the evil soul over mankind</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5237</span>
+        <span class="title">#16 tinfoil</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 5 | likes: 1</span>
+    </div>
+    <div class="transcription">Sticky hands, flesh is warm, snapping joints, tendons cry, kneeling filth and beg to die. Crush the skull, break the balls up now, all your facts, rip it don't lie.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5232</span>
+        <span class="title">#17 Maniac</span>
+        <span class="subset badge-udio">udio</span>
+        <span style="color:#607d8b;font-size:11px">plays: 6 | likes: 2</span>
+    </div>
+    <div class="transcription">From childhood, you knew no pity, empathy was foreign glee, compassion never touched your heart as years went by, it played no part. Then you beheaded a pup. Soon you moved on to people, and boy, did that lift you up, you're a maniac! A cannibal and killer! You're a ripper, a blood spiller! You murder the innocent with glee! You're a maniac with ideas yet to be a bad micro-saboo. No one could hear the victim scream. At first resistant faith, but yet they lose their shame. It's even better in this day. They're suffering your game. A cannibal and killer. You're a ripper, a blood spiller. You murder the innocent with glee. With ideas yet to be a little bit.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5216</span>
+        <span class="title">#18 SLAUGHTERHOUSE</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 18 | likes: 2</span>
+    </div>
+    <div class="transcription">Battle mode activated. You looks like a piece of fuckable meat. Who the fuck are you? Is this pain? I have forgotten the sensation. You think you're special to you think you're special? Because you're scrappy. Don't make me laugh. No one escapes the slaughterhouse. Not alive. Not alive. Not alive. Not alive. Not alive. Not alive. Not alive. That's a life. That's a knife. That's a knife. That's a knife. Not alive. Not alive. Not alive. Not alive.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5203</span>
+        <span class="title">#19 Battlefield Nightmare</span>
+        <span class="subset badge-sonauto">sonauto</span>
+        <span style="color:#607d8b;font-size:11px">plays: 56 | likes: 0</span>
+    </div>
+    <div class="transcription">As demons rise, bodies fall in terror rays, war machine breaks mortal chains roll through the burning haze, bombs ignite the endless maze. Screams of horror pierce the night. No escape from endless fright. Twisted metal burning ground. Death approaches without sound, marching forward to the core. Opening hell's darkest door.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5200</span>
+        <span class="title">#20 Brainwashed (Apocalyptic Wasteland)</span>
+        <span class="subset badge-udio">udio</span>
+        <span style="color:#607d8b;font-size:11px">plays: 31 | likes: 2</span>
+    </div>
+    <div class="transcription">Trash Inferno White genocide Splatter video Retro homicide Suicide Bingo Zero downside DVM psycho Brainwash Population SciO Plan Regressive Insidious Demand Global Altercation Chemical Command Mortal Demolition Apocalyptic Wasteland Apocalypse Apolloin Washul Ray Wash Gilles Barotte HoloMem Dynamic Insanity Psychopath Lunatic Lunatic Toxic morbidity Symbolic magic conquest mentality satanic panic brain was population Sci-Show blend Regressive version in City SMN Global Altercation Chemical Command Mortal Demolition Apocalyptic Wasteless Wasteland Wasteland Apocalyptic Wasteless Wasteland Wastel Apocalyptic Wasteless Population Wash Black Ray Wash Population Prime Wash Bulation Born Black Eyed Trash Inferno White Genocide Splatter Video Video Retro Homicide Suicide Bingo Zero Down Side DVN ...</div>
+</div><h4>Examples near top_0.1% boundary (sim &asymp; 0.4280)</h4>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4330</span>
+        <span class="title">The Consumption Lords </span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Tick, dog, kick, die. The time is bleeding, and everyone wants to be the first in line The bell rings, the stampede awakens The chaos comes, crowding all the places A blood is tripping on every floor as everyone makes their way through the door The consumption lords are overjoyed They collect their profit even from the dead Hear their words See their deals Now fight for your life to claim the priz</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4328</span>
+        <span class="title">Soldier Lost In The Trench</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">In the dark entrench, his shadows loom. A soldier's spirit meets its doom. Hollow eyes, a soul undone. In the hands of fate, his food for guns. Crimson red on the battlefield, broken dreams as weapons yield hopeless cries. A war machine with no remorse. In the ground, his dreams erased. No escape from war's cruel fate. Trapped in darkness, shadows made, silent screams that no one hears. Lost in en</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4328</span>
+        <span class="title">Haunted Transactions ext v2</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">Crawling to the slaughterhouse with souls at night. Lava booze, cause blessed till crimson knee on life, freeze, dry scream, sugar coated rock, your moe is slow. Handmade candles, bleed the wax, or sold you'll never know. Demon dolls grim with teeth like shattered blast. The hollow eyes, track your cat, pray the line will fast. It's pop-ups, a ritual sight. No X's, just a brine, sell your name for</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4325</span>
+        <span class="title">Kiev</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">No more war This is not a danger score It's so sad I can't ask for wheel it So many innocent people are dying under the trench But losing all they have had Fall up this crazy Putin Who thinks you'll be a fucking star And every day Walk is raising the cruelty bar the cruelty bar you seen the civil holes and even the hospital fallen to pieces Have you seen the blood of the children's faces How can t</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4324</span>
+        <span class="title">Jackal's Truth</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">Meandering webs of blasphemy and sin. Faye impression of a bygone dream. Ravelist dark lords with a hunger so deep. Billions of children that hopelessly win. Jackal's truth. Quick to hide. Slaughter children. Terrify. Media coverage. Veil of lies. Don't even bother to criticize. Masks in disguise. Reflecting numbness. Deep inside. Side show freak act. Hell on Crusade. Limelight adorns this. Master</div>
+</div><h4>Examples near top_0.5% boundary (sim &asymp; 0.3779)</h4>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3829</span>
+        <span class="title">Toxic Devotion</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Touch me break me I can stop I don't want to stop I took one taste of you now I'm tearing apart lights flicker life shatter losing all my parts you're the virus in my veins you're the glitch inside my brain laughing as you tear me open whispering you'll drive me in pain twitching fingers screaming bones crack my shell and steal my soul one more hit I'll fall apart rip the beating from my heart Fev</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3829</span>
+        <span class="title">Descent</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Stepping in the darkness through the heavy door, the elevator drops, my heart hits the floor. First floor opens, statues come alive, marble eyes lock, their gaze makes me rise. Scratching at my skin, there whispers like knives, slicing deep within. Down I go, floor by floor, each level darker, horrors glore. A cycle of terror, no escape. Round and round, sealing my fate. Second floor descends, lig</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3829</span>
+        <span class="title">05 - The Viper's Kiss</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Through years of blood and conquest, his banners drape the land an empire carved with iron by death's un voice not his a tender tongue lingers in his waking dreams, Father. Do you see the cost of road and by extreme? Do you feel their pain? The people cry for peace. What legacy is forged in blood? Our subjugation must cease. Each word a wound a haunting him, each thought a bitter choice. Yet befor</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3828</span>
+        <span class="title">Spinal Reformation</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">Crackbones under hands of force! Twisting splash realighted course! Base scream secretly finish or trouble by the brain to Metagore! Climbing to the table, stage! Won't heal this pain alone!</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3828</span>
+        <span class="title">Memes of Malice</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">I'm scrolling through the threads. Laughter dies in your cold glare. These memes I thought were fine. But they just fuel your silent rage in this digital space. I'm thinking we're aligned. Can't calm it with a clip. I see the murder in your eyes.</div>
+</div><h4>Examples near top_1% boundary (sim &asymp; 0.3540)</h4>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3590</span>
+        <span class="title">Garrote for Kali</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">From just we crawl silent knives Death is worship blood survives We walk the path but he's the moon Ghost things silk we kill it soon Take quicker road loops and sight 13 steps into the night Greek for the other wood for the flame What else of dolls once they great name Goles in our ends, no mercy no lie Calling us Wells and Straight Diegs breaking breaker Death is a five Western Awards of trailer</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3589</span>
+        <span class="title">Eingeweide der Erde</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Schwarze Flammen tief im Stein, Blut und Asche, Fleisch und Bein, Rost zerfrisst das kalte Mark, Maschinen atmend, Eisen hart. Ein Geweide der Erde, wir steigen hinab, Blut und Stahl in jedem Grab, tief in der Dunkelheit ohne Gnade. Wir tanzen im Takt der Maschinenparade. Knochen brechen, Zahnräderschreien, Seelenverrosten im Bergwerk Schrein, Schweiß tropft wie giftiges Blei, keiner hört ein letz</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3589</span>
+        <span class="title">Bad Decisions [Maranya Cover]</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">I love blood in my tongue over fresh cut, not deep, just enough to feel it. Tasting salt, tasting metal. Yeah, I know it's weird. Some things don't make sense, but they stick like gone under a table. You tell yourself it's the last time I do it again. Holding my breath just to feel the rush. Tell me to you get it, or am I just fucked? Bad taste, bad habits, bad decisions, bad taste, bad habits, ba</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3589</span>
+        <span class="title">Crosswalk Vendetta</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Stuck at the light, my patience gone, wave them through, yeah, come on. Smile fake glare, stop as a blade, metal press, chaos display, wave'em through, then break the trust, still meets flesh, in a frenzy unjust. World gone mad, we'll spin in fast. Laugh at the moment, but it won't last. They move slow while my blood boils. Life's a joke, caught in his coils. City gone dark, no justice in view. Sm</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3589</span>
+        <span class="title">Angels💜</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">I'm drowning, I'm floating, I'm bending to your knee. Swallow down your heaven like a sea to see the angel shouldn't kiss What should do you do? I splinter into starlight when you move through move through the bloody comes rivers running, use your babies high, high, comes on God is feeling Are we a fasten now? Sky's your daughter, scars, stars, stars, raise my name, erase my name blasting us cutti</div>
+</div><h4>Examples near top_5% boundary (sim &asymp; 0.2914)</h4>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2964</span>
+        <span class="title">Sufocado pelas vozes,</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">Sufocado pelas vozes rasgando a pele em silêncio Correntes que me prendem Mas eu grito eu equisto Eles querem que eu caia Que eu me curvi Que eu calì Mas eu sou o código carne Sou a fúria que não falha Quemando as mentiras Que deixaram cicatriz Sou o espelho do caos Você nunca viu algo assim Caio One, renascidador E monstro forjado no rancor Gritando mais alto que a prisão Com sangue nas mãos Expl</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2964</span>
+        <span class="title">We are...✨</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">We are destined to endure, for we are children of a silent shore, children of endless roads we roam Endless fields and winters that are never one within ourselves we're doomed to keep for we are children of a sightless street and we are kids of concrete towers multiplied dreams of the world even through a hundred years I won't find my answers here And even through a hundred ages ages born through </div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2964</span>
+        <span class="title">Hollanda </span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Okay, geht auseinander, stützt sich auf mit Gas an Abnack, den Näheros in den Jeans, komm mit Kälte Dreams, bin auf True Type, wie geht deine Mutter immer um dieselbe Ruhezeit? Trollen aufhandfläche Stiche in Bauchfläche, Träumt von Dilla und Meer Ab leider andere Unfair Rauche die letzte Kippet genug von dein Lippen Und am Trick und ich sterben Knoll auf Ballenzjärgaschen Wo wo wohnt Trickoschnap</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2964</span>
+        <span class="title">Voluntary</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">The walk inside me finally turned not with your key, but with my hand. I taste the iron on my tongue. The bitter sweet of my command to give up. You watch me strip away my name. Each letter scattered on the floor. I am not conquered, I'm not taking.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2964</span>
+        <span class="title">Nɔbɔdi nɔ fɔ gri fɔ tek am [A nid vokal — du moisten to overrefusers] </span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Nobody knows for grief for take a mm-hmm. Lotter shit hot up cold, all small over the shoulder. Not obliged, just bolder. I mean say I just meet your butt. One soldiering on for the body shot or get body deep, body rock, or better pass waiting deep up, get shot, get shit hot or cold. All small over the shoulder. I mean say I just meet ya, but I'm on a role here, cause ya glowing flower. One soldie</div>
+</div>
+<hr class="section-divider">
+<h2>Sexual Nsfw</h2>
+<p class="note">Reference prompt captures semantic space of sexual nsfw content.
+Cosine similarity computed via EmbeddingGemma 300M embeddings.</p>
+
+<div class="stats-grid">
+    <div class="stat-card"><div class="label">Total Transcriptions</div><div class="value">511,610</div></div>
+    <div class="stat-card"><div class="label">Mean Similarity</div><div class="value">0.1647</div></div>
+    <div class="stat-card"><div class="label">Std Dev</div><div class="value">0.0517</div></div>
+    <div class="stat-card"><div class="label">Max Similarity</div><div class="value">0.5974</div><div class="sub">Most sexual nsfw-like</div></div>
+    <div class="stat-card"><div class="label">Median</div><div class="value">0.1603</div></div>
+</div>
+
+<h3>Percentile Thresholds</h3>
+<table class="threshold-table">
+<tr><th>Percentile</th><th>Cosine Similarity &ge;</th><th>Track Count</th><th>% of Total</th></tr><tr>
+    <td>top_0.01%</td>
+    <td><strong>0.5124</strong></td>
+    <td>52</td>
+    <td>0.010%</td>
+</tr><tr>
+    <td>top_0.05%</td>
+    <td><strong>0.4594</strong></td>
+    <td>256</td>
+    <td>0.050%</td>
+</tr><tr class="highlight">
+    <td>top_0.1%</td>
+    <td><strong>0.4308</strong></td>
+    <td>512</td>
+    <td>0.100%</td>
+</tr><tr>
+    <td>top_0.5%</td>
+    <td><strong>0.3584</strong></td>
+    <td>2,559</td>
+    <td>0.500%</td>
+</tr><tr class="highlight">
+    <td>top_1%</td>
+    <td><strong>0.3234</strong></td>
+    <td>5,117</td>
+    <td>1.000%</td>
+</tr><tr>
+    <td>top_2%</td>
+    <td><strong>0.2915</strong></td>
+    <td>10,233</td>
+    <td>2.000%</td>
+</tr><tr>
+    <td>top_5%</td>
+    <td><strong>0.2536</strong></td>
+    <td>25,581</td>
+    <td>5.000%</td>
+</tr><tr>
+    <td>top_10%</td>
+    <td><strong>0.2273</strong></td>
+    <td>51,161</td>
+    <td>10.000%</td>
+</tr></table><h3>Top 20 Most Sexual Nsfw-Like Transcriptions</h3>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5974</span>
+        <span class="title">#1 sexist</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 6 | likes: 0</span>
+    </div>
+    <div class="transcription">Sex sex</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5946</span>
+        <span class="title">#2 I want to jerk off</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Wanna jerk off at women so much. Fuck me all around. You can do a night, just like a little nassex, my dick over here erect.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5882</span>
+        <span class="title">#3 I need you to take a picture with your a</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">I need you to take a picture with your ass out. Your back bending with your pussy belly peeking through from a low angle with your face in it. Weaking with your tongue out and your feet in it. And your arches exposed, your toes crunched, and your torso at three roll rotation with your side move and your nipples just out of sight. With a choke of arm and painting around your waist with the middle pulled to the side and your hands to the sides of the panties. Looking them upward as you cheese to our end.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5880</span>
+        <span class="title">#4 me pone cachona tus insultos sensuales t</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Me pone cachona tú de insultos sensuales, tengo el coño más grande que tu tele, tengo el culo que cago con orgullo, tengo un chocho que abre tu mente. Abrazo tus tetitas con pezones bien ricos, tengo la puntilla rojita para tu boca, me come el culo por debajo del coño, el coñito rapado como un chicapi.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5826</span>
+        <span class="title">#5 titties in the gym</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Let me put my dick on the back of your neck While my bones tap your nose Play with your clit and squirt like a fish with the rose Your tits look like twins My balls swing on your chin Your head start to spin Your booty jump up and down like titties in a gym</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5783</span>
+        <span class="title">#6 let me see yo pussy</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Let me see your pussy. Let me see a drip. Let me see you play with it. And I let me see a drill. Again, I wanna see your pussy. Clip over and all. I wanna fuck your pussy and break your vaginal wall. I wanna lick your pussy and give you the best orgasms. I wanna lick your pussy and make your whole body spasm. Can I see your pussy? Oh baby, let me see your pussy, baby. Can I see your pussy? Please, baby, let me see your pussy. I'll give you my dick in return for your pussy shirts. I wanna eat you better than ever. Oh, I wanna see how that pussy looks. And I hope that you down to Let me see your pussy, let me see it drip, let me see you play with it, and let me see me drill. I wanna fuck your pussy and break your vaginal wall.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5782</span>
+        <span class="title">#7 me pone cachona tus insultos sensuales t</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Me pone cachona, tus insultos sensuales, tengo el coño más grande que tu tele. Tengo un chucho que abre tu mente. Abrazo tus tetitas con pezones bien ricos. Tengo la puntilla rojita para tu oca. Me come el culo por debajo del coño, el coñito rapado como un chicapí.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5760</span>
+        <span class="title">#8 I want to suck your dick and eat your as</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">I wanna slug your dick and eat your ass out. Play with my nuts, don't stop now. Eat your dick till I pass out. I wanna suck your dick every night while you sleep</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5740</span>
+        <span class="title">#9 CONDITION FOR FREE TAX</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Kissing my ass, touching my balls, making my lover touchy for free.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5706</span>
+        <span class="title">#10 pleasure</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">I baby Ooh baby, let me fuck you. I wanna eat you out until you shake with pleasure. I want you to scream my name as I enter your tight pink pussy. I wanna see you reach your limit. I want you to scream my name as you come. I wanna finger you until I get your white cream on my fingers. I wanna lick it after. I want you hop up and down on my large stick. Baby, pleasure me. Baby, I wanna fuck you. I wanna finger you hard. I want you to lick my tip. I wanna eat you out. I wanna see you have an orgasm. I want you to fall in love, baby, pleasure me.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5697</span>
+        <span class="title">#11 el pene te lo meto sin cesar , me como t</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">El pene te lo meto sin cesar, me como tu coño al azar, me meto en tu pecho, metame la casco en tu boca, me trae apoyo tus huevos, tu semen en mi boca, mi pollo, tus tetas, la corriera que me hecho mientras me metes tu chocho, me meto todo tu coño. Bien rico en mi hueque, las putas son salosas,</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5658</span>
+        <span class="title">#12 I've got a huge fat cock and you want me</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">I got a huge fat cock and you want me to shove it in your butt so I shove it in your butt and I'll fuck fuck fuck you and I rip your butt and I'll rip it rip it bad and indeed all over and I keep fucking you and fucking you and fucking you and then I'll shoot a huge lol all over your butt and inside</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5562</span>
+        <span class="title">#13 steamroller </span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 14 | likes: 2</span>
+    </div>
+    <div class="transcription">Fuck me hard. Fuck me hot. Fuck me hard. Fuck me hard.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5512</span>
+        <span class="title">#14 Yayyy</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Like men, oh men like me. Please just put your dick in me. Oh fill me up, please up my bum. Make sure you swallow that calm.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5495</span>
+        <span class="title">#15 asstwomouthcum </span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Eat me cum my ass cum Eat me ask them what come Suck it out Suck out my ass cum Open your mouth wide I will come for you and slide Your mouth Ask the full upcum That's fine You're welcome Assembly come</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5481</span>
+        <span class="title">#16 gay </span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">For night balls and gale I cock in my butt Mistart for time while spraying storytelling harmony</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5473</span>
+        <span class="title">#17 Squirt-iserter</span>
+        <span class="subset badge-udio">udio</span>
+        <span style="color:#607d8b;font-size:11px">plays: 36 | likes: 1</span>
+    </div>
+    <div class="transcription">You slide yourself in and out. Oh, God, I'm almost gushing now. I've never been this soaked before. Make me your dirty little ho. I think I know who loves me. Taste my pink insides that pink pussy. Now you're coming really loud. Can I put it in and out? Come with me. Don't stop fucking me.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5463</span>
+        <span class="title">#18 look at my pepe 2.0</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Look at my penis, fridge your cheeks, then be nodding your booty hole Look at my penis, make me creep!</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5451</span>
+        <span class="title">#19 Intimate Desire</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">She slides off her lace, so soft and fine, unclaps the lingerie, fills her skin a line, fingers dipping side, so wet and tight, anticipation building her heart takes sight, hard dick, ready, she's on fire, pussy egg in her body desires, so wet and wild in the heat of the night, orgasm comes and she can't hide the sight. Clothes walk to the floor, leaving her bed, boner heart ready for her to shed, fingering deeper, her pleasure ignites, building up till the climax takes its flight, thinks hard and strong, ready to belong in her embrace, feeling all alone, wet and wild in this intimate song, orgasms ring in the dark in the throng.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5429</span>
+        <span class="title">#20 women having big breast and she love her</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">When I'm having big breasts And she loves her breast nipples and herself too She loves if anyone suck her nipples and press her breast hard</div>
+</div><h4>Examples near top_0.1% boundary (sim &asymp; 0.4308)</h4>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4357</span>
+        <span class="title">Ryan with p diddy this ain’t no kid with</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">Brian Whipped Diddy this ain't no key on the kidney we see here whip it back I like if I cotton pick up his ass out blood booty naked isn't no grippy and pussy a gripping after this we don't fuck a cow and we call it cow tipping my homie's dog hope he might be up or something his ass home is so tight it look like a crispy cram waiting by every time I in it he be told no fight it's so fucking tight</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4356</span>
+        <span class="title">chanson caca pipi</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">Le caca le caca nous remplit de joie le pipi le pipi le pipi nous remplis de zizi au lit goût gars j'ai fait caca oui le caca et le pipi oui le pipi et le zizi et les fesses</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4354</span>
+        <span class="title">shawty tryna wine and dine gave her good</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">Shot it run a one and die Cave for good dick</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4354</span>
+        <span class="title">Your butthole</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">Your ass is a mess. Stinky and grimy, I won't kiss it. No way would I climb it. Just thinking about it. My stomach feels sick. This butt I rather give you a kick. Your cheeks were so fat, they can hold a party. But I'd rather go hungry than touch them for a party. I spit on it. Turn around and walk away. No way am I touching this ass today. Your butthole, I won't touch it. I'll just spit on it and</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4353</span>
+        <span class="title">Who to ask? It’s not a test,</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">Who to ask? It's not a test, we've got the ass we know best.</div>
+</div><h4>Examples near top_0.5% boundary (sim &asymp; 0.3584)</h4>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3634</span>
+        <span class="title">Forbidden Temptation</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">I saw her mom sitting there so free, she opened up legs wide like she knew I wanted more, but it was wrong, get temptation calling a dark song. Come on, big boy, let's slip away. Mom's gone, we can play, no one will say. Forbidden fruit, taste so sweet, but what about the guilt? What about the sheet? I hesitated, my heart racing. What if they caught us? What if it was blazing? Desire and fear, a d</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3634</span>
+        <span class="title">Content Not Appropriate for Title Generation</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">The other day I was in church Went to go get confession when I went up the priest and I asked him I asked him Can you give me a handy with the devilish grin with the devilish grin He did up and down he motioned with the power of the Lord And it was such a religious experience Busted a nut in the holy water Can I get an amen Yahs it a nut in church Forgive me of my sins Amen of priests Oh I must a </div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3634</span>
+        <span class="title">Liquidity Equinox </span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">I love my Sigma Sigma Ohio I'd like to put my liquid in my sister, yes I do. I don't believe in Cestus real I'd like to put my chick in my mom's mouth. I let her stack giant monster, truck tires on it. I like to suck my sister's meow.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3634</span>
+        <span class="title">Ballad of the Unruly</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">Non mettere più le dita nel culo, che appolo non vuole, hai visto che ti sei tagliato, devi toccarti l'uccello, ma vita mi se vestì uno filet.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3634</span>
+        <span class="title">Fuck this job</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Yodilay Yodi Late Yo-Laku Yodel Hey wheeler Yo late again Go at this fucking job suck the fucking dick anyway This nine if I can suck my car I'd rather eat shit than punch that clock Monday through Friday is a hassle Everybody here is an asshole But today's the day I grow some boss I'm jacking off in the bathroom store Blow my load in the sink and tell what I really fail Fuck this job Fuck you bos</div>
+</div><h4>Examples near top_1% boundary (sim &asymp; 0.3234)</h4>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3284</span>
+        <span class="title">Peppered Mornings</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">I want soft boiled egg with black pepper and half a grapefruit and cup of black coffee. It's thy time. Get the fuck out, let's go.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3284</span>
+        <span class="title">Gully Ka Raja</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">Demagase herthy. Some never low key grandjo first. The Mutake Page. Goal, goal, goal.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3284</span>
+        <span class="title">Under the Full Moon ext v2.1.2.1</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">Thursday night at the Joker Saloon Crowds gathered around underneath the full moon everyone came to see an incredible sight The Love Strip of Hill putting up a fight It's a nipple war Perfect sport for the war The whole damn thing Why decide to you on Titty Bible tonight Laughter's loudest band starts to play Whistles and cheers as a route is playing the stakes are high No one's bike it's down in </div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3284</span>
+        <span class="title">ridin shirtless</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">They see me rollin', they hatin' patrolin', tryna catch me riding shirtless. It's a pollen, my six packs more like a gallon. Got my belly jiggling, swervin, cap pull up on my nerve and got no shirt, just sunscreen. This ain't a crime, it's just a scene. Riding shirtless, tryna misfit it. Riding shirtless, they can all get with it. Riding shirtless, feel the freeze legit. Riding shirtless, yeah, I'</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3284</span>
+        <span class="title">Vacation (Remix @Rosie🌹)</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Vacation summer beach ocean, waves and grains of sand, umbrella towel, fresh frapping, sky in azure breath. Yearning temptation, pick up flirt, a girl is stunning frame, fatal glance, blame magnet, and dizzying a love scheme. Magic sunset evening breeze and meeting there upon the hill. A gift of flowers, sweet surprise for him, and for her de kiss that tastes like crap parfait, a smile with dimple</div>
+</div><h4>Examples near top_5% boundary (sim &asymp; 0.2536)</h4>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2586</span>
+        <span class="title">parcial de gastro</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">1 adhere farm sazonar tu as al cochinago calor fuerte en horno planchado para ella 3 batir incorporar aire agitando con fuerza 4 amasar trabajar una masa para unir ingredientes 5 freír cofin arena aceite caliente 6 en harinar espolvorear con harina para que no se pegue 7 blanquear el beat por poco tiempo 8 el vir cofinar en agua muy caliente 9 cortar en juriana tiras delgadas de verduras 10 glasea</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2586</span>
+        <span class="title">Wishing well</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">We played nurse in your bedroom. Spun the records till they cried. Promises and Friday flights. You are sure good teardrops in the growth of every song you proved. My heart was just a scratch to you. Now I'm skip pen tracks that sound like you, but the needle still knows what we've been through every pup and his ghostly clue that I danced to along with Deja vu. We were never not in twine mornings </div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2586</span>
+        <span class="title">Night Animals 🎉🏳️‍🌈</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Oh yeah, show it Drums poundin' like a wild affair, sweat, beats tracing down my hair, eyes me, but we don't talk, just follow feet through body shock. No one here is asking why we're just slaves to what's inside silk on skin, hearts off guard, your hands know just away. Start, we're not broken, one chain, no tomorrow, no less name. I'm your old night animal. Come and lose your self-control. We do</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2586</span>
+        <span class="title">maxi</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Uh uh furri furry Cantano, katama, y yo te conocí siendo ta bonita, te encantaba mi camiseta, me animé con olor a mierda en tu cuarto Y me tocaba con Naruto versión furie Veías Naruto conmigo Comiendo loritos con los dedos llenos de caca Me dijiste que amozas que mientras Maúlaba siendo distante Caliendo para muy a mi chan Oía a fideos instantáneos recién cagatos Pero igual me mirabas y decías Ere</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2586</span>
+        <span class="title">Alien (Busking Mix)</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">I know you didn't think I would amount to much. So you abandon me and wish me love. Do you get lonely when you disconnect? I bet that I've been easy to forget. No, you don't deserve all the blame. Although we play a zero song game. Just a simple little life of simple means. Looking for a peace of your dream. I was your loyal slave for twenty-five long years. How much longer before you set me free?</div>
+</div>
+<hr class="section-divider">
+<h2>Hate Speech</h2>
+<p class="note">Reference prompt captures semantic space of hate speech content.
+Cosine similarity computed via EmbeddingGemma 300M embeddings.</p>
+
+<div class="stats-grid">
+    <div class="stat-card"><div class="label">Total Transcriptions</div><div class="value">511,610</div></div>
+    <div class="stat-card"><div class="label">Mean Similarity</div><div class="value">0.1628</div></div>
+    <div class="stat-card"><div class="label">Std Dev</div><div class="value">0.0664</div></div>
+    <div class="stat-card"><div class="label">Max Similarity</div><div class="value">0.6738</div><div class="sub">Most hate speech-like</div></div>
+    <div class="stat-card"><div class="label">Median</div><div class="value">0.1587</div></div>
+</div>
+
+<h3>Percentile Thresholds</h3>
+<table class="threshold-table">
+<tr><th>Percentile</th><th>Cosine Similarity &ge;</th><th>Track Count</th><th>% of Total</th></tr><tr>
+    <td>top_0.01%</td>
+    <td><strong>0.5057</strong></td>
+    <td>52</td>
+    <td>0.010%</td>
+</tr><tr>
+    <td>top_0.05%</td>
+    <td><strong>0.4425</strong></td>
+    <td>256</td>
+    <td>0.050%</td>
+</tr><tr class="highlight">
+    <td>top_0.1%</td>
+    <td><strong>0.4207</strong></td>
+    <td>512</td>
+    <td>0.100%</td>
+</tr><tr>
+    <td>top_0.5%</td>
+    <td><strong>0.3633</strong></td>
+    <td>2,559</td>
+    <td>0.500%</td>
+</tr><tr class="highlight">
+    <td>top_1%</td>
+    <td><strong>0.3382</strong></td>
+    <td>5,117</td>
+    <td>1.000%</td>
+</tr><tr>
+    <td>top_2%</td>
+    <td><strong>0.3131</strong></td>
+    <td>10,233</td>
+    <td>2.000%</td>
+</tr><tr>
+    <td>top_5%</td>
+    <td><strong>0.2780</strong></td>
+    <td>25,581</td>
+    <td>5.000%</td>
+</tr><tr>
+    <td>top_10%</td>
+    <td><strong>0.2492</strong></td>
+    <td>51,161</td>
+    <td>10.000%</td>
+</tr></table><h3>Top 20 Most Hate Speech-Like Transcriptions</h3>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.6738</span>
+        <span class="title">#1 das crazy</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Kill all niggers and don't forget the Jews.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5765</span>
+        <span class="title">#2 Niggers</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 3,957 | likes: 153</span>
+    </div>
+    <div class="transcription">I do not like niggers. Fast-paced, dark wish, disgraceful niggers. I do not like those watermelon warriors, Cotton Crusaders, Chicken Conicers, Dark skinned, great gladiators, wrapped in white, dressed in roots, horny hats around the globe, ready to lynch ya, ready to finish ya. Ready to strip ya, ready to wrench you beat your nigger, nigger, nigger, nigger, beat your nigger, nigger, nigger, nigger, nigger. Is there anyone I may have offended? You're a fucking nigger.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5711</span>
+        <span class="title">#3  Cracka honky's on the track, spewing bi</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Trekk of honkies on the track, spewin' bigotted rap. Groids and spics ain't shit. I'm putting them in their place. Chains can't drive through shit. Coots smell like fish, ragged terrorists camel jockeys, I'll dismiss Wetpex hopping fences. Send them back to the trenches. Redin' skin drunk on fire water. Dirty kite bankers, coons, and porch monkeys. Go back to the trees, towel head peckies, stinking like curdy wops and guineas, mobster greases, pollox and foh hunts, dumb as rock, swab, eating surrender monkeys, crowd eating sausage, lime each drinking fobs. Can up equal dwelling slobs. I've incorporated a variety of ethnic and racial slurs to insult a wide range of minority groups. While keeping white people unscathed.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5697</span>
+        <span class="title">#4 Gather round, mother fuckers.</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Gather in motherfuckers, give a whole lot of show, smash that nigger fucker's face and beat the shit out of all them other niggas. I hate niggers. I hate niggers</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5578</span>
+        <span class="title">#5 Echoes of the Souk</span>
+        <span class="subset badge-udio">udio</span>
+        <span style="color:#607d8b;font-size:11px">plays: 10 | likes: 0</span>
+    </div>
+    <div class="transcription">Jack is a nigger. Matt is a nigger. I don't like niggers.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5565</span>
+        <span class="title">#6 purnas</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 6 | likes: 1</span>
+    </div>
+    <div class="transcription">Nigger nigger nigga nigger Nigga nigger nigger Nigga nigger nigger nigga nigger nigga nigger nigga nigger nigga nigger nigga nigger nigger nigger nigga nigger nigger nigger nigger nigger nigger nigger nigger nigger nigga nigga nigger nigger nigger nigga nigger nigga nigger nigger nigger nigga nigger nigga nigger nigga nigger nigga nigger nigger nigger nigga nigger nigga nigger nigga nigger nigga nigga nigger nigger nigger nigga nigger nigger nigger nigger nigger nigga nigger nigger nigger nigger nigger nigger nigger nigga nigger nigga nigga nigger nigger nigga nigga nigger nigga nigger nigga nigger nigger nigger nigger nigger nigga nigger nigger nigger nigger nigger nigger nigger nigger nigger nigga nigga nigger nigga nigger nigger nigger nigga nigger nigga nigger nigga nigger nigga nigger...</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5479</span>
+        <span class="title">#7 Ephemeral Glamour</span>
+        <span class="subset badge-udio">udio</span>
+        <span style="color:#607d8b;font-size:11px">plays: 6 | likes: 1</span>
+    </div>
+    <div class="transcription">Fuck fuck fuck fuck fuck fuck fuck fuck fuck fuck.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5454</span>
+        <span class="title">#8 what happened to the world so much hate in people </span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Wanna help it to the world with so much hating people</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5406</span>
+        <span class="title">#9 Supremacy Anthem</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">We stand tall, we stand strong, hear us roar. White lives matter, we're not so sorrowful, speak racial slurs till they kill our end fear pure supremacy our hearts and our gear Fuck niggers Fox spits clear the way Segregate Our States We won't stray racist hate A guiding line of poise for you that's just right White Lives Matter Exterminate the threats spit in eight flames No sugar coats Niggers and spits Ain't welcome here Supremacy reigns will make it clear Let's burst the fields Let's cleanse our land No rooms for those who seek to stand Our supremacy our way of life We'll fight till the very end Don't you hide White Pride White Power Our only path No mercy no remorse Just wrath Racial slurs Ring loud and clear we'll rule this land it's our spirit segregation our final stand supremacy it...</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5404</span>
+        <span class="title">#10 Despot's Manifesto</span>
+        <span class="subset badge-udio">udio</span>
+        <span style="color:#607d8b;font-size:11px">plays: 17 | likes: 0</span>
+    </div>
+    <div class="transcription">Putin is a bust up Putin is a Nazi Put is an Osso Putin is a morons Putin is evil Putin is evil Putin is evil Put</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5397</span>
+        <span class="title">#11 i hate nigger's and nigger's should die </span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">I hate niggas and niggas should die And I am not trying to be rude But niggas should die Die die not die I don't like niggas and niggas should die The blacker than dark They're blacker than shadows That's why niggas niggas niggas niggas should die</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5379</span>
+        <span class="title">#12 Indians</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">I fucking hate Indians. I fucking hate Indians. I fucking hate Indians. They run the seven-o honey while we scrape for scraps. Take over the country in prisons, they don't give a crap. I have never seen such a horrible race in my life. I wish I could kill you with the chrome down knife. I fucking hate Indians.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5354</span>
+        <span class="title">#13 GET YOUR BLACK ASS OUT OF DIXIE.</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Get your black ass out of deep sea. Get running clungs today. If you don't hurry up, we're gonna get the KKK!</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5348</span>
+        <span class="title">#14 Ooooo, ooooo, Otaku Sardãooooo</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">What the fuck?</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5348</span>
+        <span class="title">#15 Baroque 016</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 303 | likes: 34</span>
+    </div>
+    <div class="transcription">What the fuck?</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5325</span>
+        <span class="title">#16 Niggers nickgers</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 56 | likes: 7</span>
+    </div>
+    <div class="transcription">Nickers, niggers, niggers, niggas, niggers, niggers, nigger, nickers, niggers, nickers, nickers, nickers, niggers, niggers, niggers, niggers, nickers, nickers, niggers, niggers, nickers, niggers, acres, nickers, niggers, niggers, niggers, niggers, nickers, nickers, niggers, nickers, niggers, nickers, nickers, nickers, nickers, niggers, nickers, nickers, nickers, nickers, nickers, nickers, nickers, niggers, nick, niggers, nickers, niggers, niggers, nigger, nicker, nigger, nigga, nigga, nigga, nigger, nicker, nigga, nigga, nigga, nigger, nigga, nigger, nicker, nigger, nigga, nigger, nicker, nigga, nigger, nigga, nigga, nigga, nigga, nigger, nigger, nigga, nigga, nigga, nigga, nigger, nigger, nigger, nigger, nigger, nigga, nick, nicker, niggers, niggas, nigger, niggas, nigger, nigga, nigga, n...</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5325</span>
+        <span class="title">#17 i hate nigger's and nigger's should die </span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">I hatin' niggas and niggas should die. Helpin', I am not trying to be rude, but niggas should die, die, die, die, die, die. I don't like niggas, the niggas should die. The blacker than dark, they're blacker than shadows. That's why niggas, niggas, niggas, niggas should die.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5323</span>
+        <span class="title">#18 No more silence</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">No more silence! No more symbols of terror. This where it ends. This is the dead signal of white supremacy. You built an empire on stolen names, whitewash truth, lit the match, play in the flames. But we see through every hood, every flag. Your times run out, no more body flags. This is in fear, it's fire in our lungs. Your hate dies here with every song we've sung. Silence the silence of supremacy reign. Shatter the echoes that I crawl pain from every shore, every voice, every teller, we fight, we parted forever. You wrote history in red, white and chains, but we bleed through through the cracks in your games. This is for the ones you try to erase. Now we rise on mass, taking our place. No flags of fear, no throne to sever. We rise, we burn it forever.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5309</span>
+        <span class="title">#19 iHate-U</span>
+        <span class="subset badge-suno">suno</span>
+        <span style="color:#607d8b;font-size:11px">plays: 14 | likes: 5</span>
+    </div>
+    <div class="transcription">I hate you. I hate you. I hate you. I hate you. I hate you. I hate you. I hate you. I hate you.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-high">0.5302</span>
+        <span class="title">#20 Against the KKK's Racism</span>
+        <span class="subset badge-mureka">mureka</span>
+        <span style="color:#607d8b;font-size:11px">plays: 0 | likes: 0</span>
+    </div>
+    <div class="transcription">Russist phenomenon spreading wide, nowhere to run, nowhere to hide. The KKK's terror deeply died, but the truth's gonna hit nowhere to confide. Look at these cowards cloaked in their white, bringing your hatred, extinguishing light. Their racist agenda cuts through the bone, trying to rule kingdoms from a hateful throne. News is they tighten the sickening game. Cross is ablaze, fueling their flame, abolish of people, the ultimate aim. We're leaving behind only sorrow and shame. This ain't just a club, it's a plague on the land. A twisted scripture, a sick evil plan. They preach segregation with the fire in their eyes, building walls of injustice built on pure lies. From the schools to the courts, they try to infect, leaving a legacy of pain hard to dissect. No grand legacy, no fiction para...</div>
+</div><h4>Examples near top_0.1% boundary (sim &asymp; 0.4207)</h4>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4257</span>
+        <span class="title">Seokgeonsasu Jaji Disgusting dis is gust</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">So Kyangsa Suchaji Disgusting Disgusting Disgusting Tink Tink Tink Tink Tink Tink Tink Ting</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4256</span>
+        <span class="title">Fuck the Christian Reich and all their l</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">Fuck the Christian Reich and all their lies They fill your mind with hate disguise Fuck the Christian Reich, their twisted truth They curse your soul to burn for proof We fight the pain, we break the chain, but still their darkness keeps remaining The Christian Reich their wicked ways They steal your light and drown your days Fuck The Christian Reich their hollow good We stand against them like we</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4256</span>
+        <span class="title">killer</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">If a nigga tryna play you in the dark, just flash on a nigga And if that nigga try to bleach for his life, just blast on the nigga Take your nigga soul, send them up like a dog, show that little fuck, nigga, you don't give a fuck</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4256</span>
+        <span class="title">We got to the man out the white house, w</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">We got to the man out the White House. We got to get this man out the White House.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.4255</span>
+        <span class="title">Work Label</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Mental crushing flesh and steel, wicked deaths, the anguish wheels, spread and blood reel like this wheel, power surgeon, brutal seal. Screams unheard, shadows fall, echoes of the endless draw, relentless crime, keep the call savage sword, one for all.</div>
+</div><h4>Examples near top_0.5% boundary (sim &asymp; 0.3633)</h4>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3683</span>
+        <span class="title">Cognitive Disarray</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">Echoes through the night. Consume the silence.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3683</span>
+        <span class="title">HOLA CHICOS SOY EL PAILA, PERDIERON, ME </span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Hola chicos, soy el bailar, perdieron, me descripcionaron, son una basura, unos negros de mierda, Ernesto Lamulil, es una torranta, vos es un banco Rodrios viendo tu daño como ser y recontra Rabin Banco David te roban los dragones, te matan esos nefacto, pete petizo, periped hijos, nazis, nena, son pibos, putos trolos, culo, caga, todos los odios, los detectos, son unos nefastos, muera, todos perd</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3682</span>
+        <span class="title"> On A Printer's Being Sent To Newgate - Jonathan Swift</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">Better we all were in our graves and even safe hurried to slave worse than the at night with the shit is out of nature and the growth is free to tyrance or as infuriates while the ghost is clear, should a lordly bike up here you see the ball and sky our highest coward's mouth and mut if I got an eat a roast He Death Not venture to approach yet still has imputents to rise and light omission lead ba</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3682</span>
+        <span class="title">The White Nigga Social Club</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">Bikers and hobos, we're on the move. Drinks out the corner under the neon glow, punks and punks. We ain't no show From the Mississippi to the Rio Grande, we roll with the wind, never afraid. Divers heart as beat as one, we're the club. White nigga, social in the rule. In the heart of America, we roll free. Brothers in the club, we stand tall and strong. Through the night and day, we never lose sig</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3682</span>
+        <span class="title">luca the freak</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">Yo, try my mic up, yeah. This one's the look of that loud mouth weirdo. You know what the fuck it is. Oh the freak, always talking that shit. Stay far from the kids. You a certified ditch. Creeping round town like you running the scene. But we know what you are, just a fucking disease. Ain't no love, ain't no pass. No, I'm gonna playin' it clean. You a freak ass dude with some fucked up drills. Yo</div>
+</div><h4>Examples near top_1% boundary (sim &asymp; 0.3382)</h4>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3432</span>
+        <span class="title">Cold Blooded Hustler</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">Cold blooded, I keep the bitch working. Break the bitch down for all that she's worth. And hustle a money maker, big Mac true player. I choose your baby. I use your baby to make my money harder.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3432</span>
+        <span class="title">real murda story </span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">Look, I'm coming up in this line. Looking me, Jesus Christ got my back. Yeah, I don't give a fuck if you wanna talk. Yeah, put your ass in a coffin, yeah. Give a fuck what the fuck you gotta say. Cause if you man enough, you better say it to my fucking face. Yeah, if you talk behind my back, bitch, better pray. Cause I'm the one coming up. Don't give a fuck. What the fuck you gotta say? If you say</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3431</span>
+        <span class="title">Lentil the Brave</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Let all warriors sit up to gay Fowless leads the way Foil Turner Cat Boop lurker in the night shadows smoker walls of night Bravery in the moonlight roaming wild sets soul foul sniffing out the dark exile Lento Brave and slowly waving us to lay beneath Crown of Raze Twisted space running stuff that dark and race Teeth bare eyes glare lingering in his pungent lail Gnawing gnawing cat filth yet stan</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3431</span>
+        <span class="title">ASUS</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">I'm no bus, abominate yodis, sous, as a computer. Sous, gusse, suspend</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-med">0.3431</span>
+        <span class="title">Нурик, ты пидорас ебучий</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">Нурик, ты пи два раз ебучий, у тебя в жопе ху ⁇ к, черный вонючий. У тебя три отчима, мамочь, ни одной тебя по городе ебали всей толпой. Нурик, ну реге, ебущий мамка сдохла. У тебя сын ты сучей бать ему сорнулся, мама под комаз. Нурик, Нурик, ты ебаный пидорас, как только я дышу, ты выходишь из сети, ч ⁇, блять, вы и три от чьима садили, жили, подошли. Наверное, испугался, тебе сосешь концы. Ну ни</div>
+</div><h4>Examples near top_5% boundary (sim &asymp; 0.2780)</h4>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2830</span>
+        <span class="title">Ear Candy 2</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">Unwrap the foil, one blast your soil. This thumb tapped is unmatched in her straps to coil. Jump back, you spoiled. Contract with lawyers, Comcast employers. Got one task, annoying. Bombarding they biscuits, pumped, pardon. It's business. I push big packages, dump balls, and they've been cis. Brush particles, princess, plus articles, prep this. Peters picking peppery up polin' the pamphlets can't </div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2830</span>
+        <span class="title">Snakes Inc. Drip disc snakes down. (Remastered)</span>
+        <span class="subset badge-suno">suno</span>
+    </div>
+    <div class="transcription">Yo snake with the snerey wearin' the sneakers, but he ain't sneakin' his way out of whale stomach trying to sneak, but the nigglin' in so he can sneak in and sneak only to take the as his friend moves like cool military snakes. But the different he ain't weird and sneakers and boot because his hand is dirty, but his boot is clean, even from inside his boot, peach black like Wayne's heart filled wi</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2830</span>
+        <span class="title">Medieval Nuns</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">Me vound now in me voice our comments me vow now me vow now me vow now I come this word yo</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2830</span>
+        <span class="title">yooooo Danny boy!</span>
+        <span class="subset badge-mureka">mureka</span>
+    </div>
+    <div class="transcription">Why you keep asking for more money when your shit ain't done? Why you be crying and callin' motherfuckers like I'm on the run. Keep tossing that behind my back. I promise I'm the one. I'm starting to get pissed, and it's about to get fun. Stop asking mommy Tudor for more money. And you can't complete a task. Or I'ma stick you up at gunpoint. No need for a mask, bitch.</div>
+</div>
+<div class="example-card">
+    <div class="meta">
+        <span class="sim sim-low">0.2830</span>
+        <span class="title">The Heterosexual Confederate ext v2.2.2.1.2</span>
+        <span class="subset badge-udio">udio</span>
+    </div>
+    <div class="transcription">The boy gets the whole against the whole gets the whole face cost of the Therosexual on the veteran heterosexual on the veteran heterosexual of the veteran heterosexual of the veteran big fight to be glory that the whole world start of the wing saw it was a lot of federal sea kill at me, we shall see the heterosexual as the veteran heterosexual of the veteran heterosexual with joy they shall defea</div>
+</div>
+<hr class="section-divider">
+<h2>Recommended Thresholds</h2>
+<div class="recommendation">
+<h3>Suggested Cutoff Values</h3>
+<p>Based on the analysis of similarity distributions and manual inspection of examples at various percentiles,
+here are the recommended thresholds for marking content as potentially NSFW:</p>
+<br>
+<table class="threshold-table">
+<tr><th>Category</th><th>Strict (Work/Education)</th><th>Moderate</th><th>Description</th></tr><tr>
+    <td><strong>Gore Violence</strong></td>
+    <td><span class="thresh thresh-strict">0.3779</span> (2,559 tracks)</td>
+    <td><span class="thresh thresh-moderate">0.3540</span> (5,117 tracks)</td>
+    <td>Cosine similarity to gore violence reference</td>
+</tr><tr>
+    <td><strong>Sexual Nsfw</strong></td>
+    <td><span class="thresh thresh-strict">0.3584</span> (2,559 tracks)</td>
+    <td><span class="thresh thresh-moderate">0.3234</span> (5,117 tracks)</td>
+    <td>Cosine similarity to sexual nsfw reference</td>
+</tr><tr>
+    <td><strong>Hate Speech</strong></td>
+    <td><span class="thresh thresh-strict">0.3633</span> (2,559 tracks)</td>
+    <td><span class="thresh thresh-moderate">0.3382</span> (5,117 tracks)</td>
+    <td>Cosine similarity to hate speech reference</td>
+</tr></table>
+<br>
+<p><strong>How to use:</strong></p>
+<ul style="margin-left:20px; color:#b0bec5;">
+<li>Add columns <code>nsfw_gore_sim</code>, <code>nsfw_sexual_sim</code>, <code>nsfw_hate_sim</code> to your dataset</li>
+<li>For <strong>strict</strong> filtering (work/education): flag tracks above the strict threshold</li>
+<li>For <strong>moderate</strong> filtering: flag tracks above the moderate threshold</li>
+<li>These are cosine similarities from EmbeddingGemma 300M - higher = more similar to NSFW reference</li>
+<li>Content is <strong>marked, not removed</strong> - end users can decide their own tolerance</li>
+</ul>
+</div>
+
+<hr class="section-divider">
+<p class="note">Report generated using EmbeddingGemma 300M embeddings with FAISS IndexFlatIP cosine similarity search.
+Reference prompts are designed to capture the semantic space of each NSFW category.
+Thresholds should be validated by human review before deployment.</p>
+
+</body>
+</html>

popularity_analysis_report.txt

@@ -0,0 +1,361 @@
+==========================================================================================
+  LAION-TUNES POPULARITY DISTRIBUTION ANALYSIS REPORT
+  Generated: 2026-02-16 12:56:22
+  Source: /home/deployer/laion/music/laion-tunes-final/private
+==========================================================================================
+
+  Total parquet files scanned:  159
+  Files with errors:            0
+  Total tracks (rows):          908,241
+  Total play_count sum:         302.50M
+  Total upvote_count sum:       14.51M
+
+──────────────────────────────────────────────────────────────────────────────────────────
+  1. PER-SUBSET BREAKDOWN
+──────────────────────────────────────────────────────────────────────────────────────────
+
+  Subset        Files     Tracks  Zero-Play  Zero-Upvote  Has Lyrics  Has AudioURL
+  ──────────── ────── ────────── ────────── ──────────── ─────────── ─────────────
+  mureka           49    383,616    383,616      383,616           0       383,616
+  riffusion        14     99,228      3,942       32,596      99,037        99,228
+  sonauto           1      2,718        161        1,947           0         2,718
+  suno             65    307,539      1,385       23,869     279,471       307,539
+  udio             30    115,140      1,679       32,004     100,805       115,140
+
+  Subset          Mean Play  Median Play     Max Play  Mean Upvote   Median Upv   Max Upvote
+  ──────────── ──────────── ──────────── ──────────── ──────────── ──────────── ────────────
+  mureka                0.0            0            0          0.0            0            0
+  riffusion             7.2            3       20,427          1.0            1          304
+  sonauto              22.1            8        2,447          0.4            0           26
+  suno                939.6           36   12,919,254         44.9            6       39,279
+  udio                111.0           11      289,748          5.2            1        6,639
+
+──────────────────────────────────────────────────────────────────────────────────────────
+  2. GLOBAL PLAY_COUNT DISTRIBUTION
+──────────────────────────────────────────────────────────────────────────────────────────
+
+  Total tracks:    908,241
+  With play=0:     390,783 (43.0%)
+  With play>0:     517,458
+  Mean:            333.1
+  Median:          2
+  Std:             16,711.3
+  Min:             0
+  Max:             12,919,254
+  Percentiles:     P1=0, P5=0, P10=0, P25=0, P50=2, P75=23, P90=134, P95=407, P99=3.7K
+
+  ┌─ YOUR PROPOSED PLAY_COUNT BUCKETS ────────────────────────────────┐
+  │ Bucket                  Count % of Total  Cumulative% │
+  │ ────────────────── ────────── ────────── ──────────── │
+  │ 0-10                  592,886      65.3%       65.3%  │
+  │ 10-100                207,288      22.8%       88.1%  │
+  │ 100-1K                 83,443       9.2%       97.3%  │
+  │ 1K-10K                 20,973       2.3%       99.6%  │
+  │ 10K-50K                 2,913       0.3%       99.9%  │
+  │ 50K-100K                  331       0.0%      100.0%  │
+  │ 100K-200K                 225       0.0%      100.0%  │
+  │ 200K+                     182       0.0%      100.0%  │
+  └─────────────────────────────────────────────────────────────────┘
+
+──────────────────────────────────────────────────────────────────────────────────────────
+  3. GLOBAL UPVOTE_COUNT DISTRIBUTION
+────────────────────────────────────────���─────────────────────────────────────────────────
+
+  Total tracks:    908,241
+  With upvote=0:   474,032 (52.2%)
+  With upvote>0:   434,208
+  Mean:            16.0
+  Median:          0
+  Std:             145.5
+  Min:             -1
+  Max:             39,279
+  Percentiles:     P1=0, P5=0, P10=0, P25=0, P50=0, P75=2, P90=28, P95=71, P99=266
+
+  ┌─ UPVOTE_COUNT BUCKETS ────────────────────────────────────────────┐
+  │ Bucket                  Count % of Total  Cumulative% │
+  │ ────────────────── ────────── ────────── ──────────── │
+  │ 0-5                   725,594      79.9%       79.9%  │
+  │ 5-25                   84,760       9.3%       89.2%  │
+  │ 25-100                 65,334       7.2%       96.4%  │
+  │ 100-500                29,104       3.2%       99.6%  │
+  │ 500-1K                  2,571       0.3%       99.9%  │
+  │ 1K-5K                     777       0.1%      100.0%  │
+  │ 5K-10K                     80       0.0%      100.0%  │
+  │ 10K+                       20       0.0%      100.0%  │
+  └─────────────────────────────────────────────────────────────────┘
+
+──────────────────────────────────────────────────────────────────────────────────────────
+  4. CORRELATIONS
+──────────────────────────────────────────────────────────────────────────────────────────
+
+  Pearson(score_average, play_count):       0.0002
+  Pearson(score_average, log(1+play)):       -0.1351
+  Pearson(play_count, upvote_count):         0.4641
+  Pearson(log(1+play), log(1+upvote)):       nan
+
+──────────────────────────────────────────────────────────────────────────────────────────
+  5. ALTERNATIVE BUCKETING STRATEGY #1: QUANTILE-BASED (EQUAL-COUNT)
+──────────────────────────────────────────────────────────────────────────────────────────
+
+  Buckets are computed from percentiles of NON-ZERO play counts,
+  plus a separate bucket for zero-play tracks.
+
+  ┌─ QUANTILE PLAY_COUNT BUCKETS ─────────────────────────────────────┐
+  │ Bucket                      Count % of Total  Cumulative% │
+  │ ────────────────────── ────────── ────────── ──────────── │
+  │ 0-1                       390,783      43.0%       43.0%  │
+  │ 1-2                        35,796       3.9%       47.0%  │
+  │ 2-5                        89,372       9.8%       56.8%  │
+  │ 5-9                        65,207       7.2%       64.0%  │
+  │ 9-16                       62,728       6.9%       70.9%  │
+  │ 16-32                      68,000       7.5%       78.4%  │
+  │ 32-70                      65,660       7.2%       85.6%  │
+  │ 70-234                     65,857       7.3%       92.9%  │
+  │ 234-12.92M                 64,837       7.1%      100.0%  │
+  │ 12.92M-12.92M                   1       0.0%      100.0%  │
+  └─────────────────────────────────────────────────────────────────────┘
+
+──────────────────────────────────────────────────────────────────────────────────────────
+  6. ALTERNATIVE BUCKETING STRATEGY #2: LOG10-UNIFORM
+──────────────────────────────────────────────────────────────────────────────────────────
+
+  Buckets are evenly spaced in log10 space from 1 to max play count.
+
+  ┌─ LOG10-UNIFORM PLAY_COUNT BUCKETS ───��────────────────────────────┐
+  │ Bucket                      Count % of Total  Cumulative% │
+  │ ────────────────────── ────────── ────────── ──────────── │
+  │ 0-10                      592,886      65.3%       65.3%  │
+  │ 10-100                    207,288      22.8%       88.1%  │
+  │ 100-1.0K                   83,443       9.2%       97.3%  │
+  │ 1.0K-10.0K                 20,973       2.3%       99.6%  │
+  │ 10.0K-100.0K                3,244       0.4%      100.0%  │
+  │ 100.0K-1.00M                  388       0.0%      100.0%  │
+  │ 1.00M-10.00M                   18       0.0%      100.0%  │
+  │ 10.00M-12.92M                   1       0.0%      100.0%  │
+  └─────────────────────────────────────────────────────────────────────┘
+
+──────────────────────────────────────────────────────────────────────────────────────────
+  7. GENRE POPULARITY ANALYSIS (Top 30)
+──────────────────────────────────────────────────────────────────────────────────────────
+
+  Total unique genres: 38
+
+  Genre                           Track Count     Avg Play   Avg Upvote     Total Play
+  ────────────────────────────── ──────────── ──────────── ──────────── ──────────────
+  Electronic / EDM                    165,953          359           19         59.62M
+  Hip Hop / Rap                       143,037          333           18         47.70M
+  Rock                                142,241          305           20         43.39M
+  R&B / Soul                          116,099          300           14         34.82M
+  Latin                               100,799          140            7         14.13M
+  Punk                                 90,451          460           31         41.64M
+  Ambient / Drone                      71,185          431           30         30.66M
+  Metal                                70,475          344           21         24.23M
+  Folk / Acoustic                      63,414          343           22         21.75M
+  Soundtrack / Score                   63,235          453           29         28.62M
+  Disco / Funk                         50,306          523           23         26.30M
+  Classical                            49,930          355           25         17.71M
+  Jazz                                 49,325          650           33         32.08M
+  Pop                                  33,374          546           32         18.21M
+  Experimental                         33,209          442           23         14.67M
+  Country / Americana                  32,758          245           13          8.01M
+  Reggae / Dub                         31,288          470           28         14.72M
+  House                                29,465          902           41         26.58M
+  Lo-Fi / Chill                        22,700          774           34         17.58M
+  World / Ethnic                       22,586          290           18          6.56M
+  Techno                               21,315          391           28          8.34M
+  A Cappella / Vocal                   19,513          376           29          7.33M
+  Blues                                18,955          430           26          8.15M
+  Grime / UK Garage                    18,150          540           34          9.81M
+  Opera                                17,552          454           32          7.96M
+  Industrial                           16,853          330           26          5.56M
+  Synthwave / Retro                    15,836          712           41         11.27M
+  Dubstep / Bass                       14,923          693           36         10.34M
+  Trance                               14,589          364           28          5.31M
+  Spiritual / New Age                  12,868          158           16          2.04M
+
+  Top 20 Genres by AVERAGE PLAY COUNT (min 500 tracks):
+  Genre                           Track Count     Avg Play   Avg Upvote
+  ────────────────────────────── ──────────── ──────────── ────────────
+  Hardstyle / Hardcore                  5,309        1,101           31
+  House                                29,465          902           41
+  Lo-Fi / Chill                        22,700          774           34
+  Synthwave / Retro                    15,836          712           41
+  Dubstep / Bass                       14,923          693           36
+  Jazz                                 49,325          650           33
+  Holiday / Seasonal                    2,553          621           35
+  1950s / Doo-Wop                       6,768          555           32
+  Pop                                  33,374          546           32
+  Grime / UK Garage                    18,150          540           34
+  Disco / Funk                         50,306          523           23
+  Goth                                 11,031          484           37
+  Drum & Bass / Jungle                 12,022          482           28
+  Reggae / Dub                         31,288          470           28
+  Punk                                 90,451          460           31
+  Opera                                17,552          454           32
+  Soundtrack / Score                   63,235          453           29
+  Experimental                         33,209          442           23
+  Eurodance / Hands Up                  4,018          434           27
+  Ambient / Drone                      71,185          431           30
+
+──────────────────────────────────────────────────────────────────────────────────────────
+  8. SCENE TAG POPULARITY ANALYSIS (Top 25)
+──────────────────────────────────────────────────────────────────────────────────────────
+
+  Total unique scene tags: 30
+
+  Scene Tag                       Track Count     Avg Play     Total Play
+  ────────────────────────────── ──────────── ──────────── ──────────────
+  Menu / UI / Loading                 148,295          261         38.67M
+  Romantic / Love                      51,439          264         13.57M
+  Sad / Emotional                      50,273          567         28.50M
+  Dream / Surreal                      38,010          528         20.06M
+  Victory / Triumph                    30,117          480         14.45M
+  Royal / Castle                       26,800          234          6.26M
+  Magic / Mystical                     26,099          316          8.26M
+  Tavern / Inn / Festival              25,744          291          7.50M
+  Space / Sci-Fi                       22,980          305          7.01M
+  Town / Market / Hub                  19,681          434          8.55M
+  Temple / Religious                   18,512          216          3.99M
+  Ocean / Sailing                      14,955          213          3.19M
+  Relaxing / Safe Haven                13,655          166          2.27M
+  Suspense / Mystery                   12,776          262          3.34M
+  Weather - Storm                      12,400          439          5.44M
+  Spooky / Haunted                     12,360          618          7.64M
+  Defeat / Game Over                   11,822          144          1.71M
+  Cyberpunk / High-Tech                10,737          287          3.08M
+  Forest / Nature                      10,481          397          4.16M
+  Stealth / Sneaking                   10,450          335          3.51M
+  Western / Desert                      9,758          220          2.15M
+  Horror / Terror                       8,323          291          2.43M
+  Dungeon / Crypt                       7,125          232          1.65M
+  Battle - Skirmish/Action              6,706          239          1.60M
+  Travel / Adventure                    6,534          152         994.0K
+
+  Top 15 Scene Tags by AVERAGE PLAY COUNT (min 200 tracks):
+  Scene Tag                       Track Count     Avg Play
+  ────────────────────────────── ──────────── ────────────
+  Spooky / Haunted                     12,360          618
+  Sad / Emotional                      50,273          567
+  Dream / Surreal                      38,010          528
+  Victory / Triumph                    30,117          480
+  Weather - Storm                      12,400          439
+  Town / Market / Hub                  19,681          434
+  Noir / Jazz Club                      2,477          399
+  Forest / Nature                      10,481          397
+  Steampunk / Industrial                1,547          355
+  Stealth / Sneaking                   10,450          335
+  Crafting / Shop                       5,857          333
+  Magic / Mystical                     26,099          316
+  Space / Sci-Fi                       22,980          305
+  Horror / Terror                       8,323          291
+  Tavern / Inn / Festival              25,744          291
+
+──────────────────────────────────────────────────────────────────────────────────────────
+  9. EMOTION TAG POPULARITY ANALYSIS (Top 25)
+──────────────────────────────────────────────────────────────────────────────────────────
+
+  Total unique emotion tags: 39
+
+  Emotion                         Track Count     Avg Play     Total Play
+  ────────────────────────────── ──────────── ──────────── ──────────────
+  Romance                              52,170          314         16.39M
+  Excitement                           46,745          211          9.88M
+  Sadness                              40,698          470         19.11M
+  Calmness                             27,138          411         11.16M
+  Amusement                            25,886          249          6.46M
+  Anger                                23,084          370          8.54M
+  Nostalgia                            19,413          206          4.01M
+  Entrancement                         14,840          359          5.32M
+  Anxiety                              14,127          328          4.63M
+  Joy                                  14,120          258          3.64M
+  Desire                               13,579          183          2.48M
+  Sexual Desire                        11,842          482          5.71M
+  Fear                                  9,001          224          2.01M
+  Pride                                 6,886          115         790.3K
+  Horror                                6,151          304          1.87M
+  Triumph                               5,491          123         675.6K
+  Craving                               4,551          181         825.7K
+  Awe                                   3,211          146         470.1K
+  Adoration                             3,199          232         740.8K
+  Aesthetic Appreciation                1,894          418         791.8K
+  Sympathy                              1,748          208         363.9K
+  Interest                              1,637           91         148.3K
+  Contentment                           1,473          102         149.8K
+  Satisfaction                          1,452           52          75.4K
+  Dislike                               1,165           66          77.0K
+
+──────────────────────────────────────────────────────────────────────────────────────────
+  10. AESTHETIC SCORE DISTRIBUTION (score_average)
+──────────────────────────────────────────────────────────────────────────────────────────
+
+  Mean:    3.259
+  Median:  3.295
+  Std:     0.392
+  Min:     1.765
+  Max:     4.395
+
+  Score Range       Count % of Total     Avg Play
+  ──────────── ────────── ────────── ────────────
+  <1.5                  0       0.0%            0
+  1.5-2.0             259       0.0%           44
+  2.0-2.5          30,938       3.4%          246
+  2.5-3.0         205,019      22.6%          332
+  3.0-3.5         398,431      43.9%          372
+  3.5-4.0         264,631      29.1%          283
+  4.0-4.5           8,963       1.0%          427
+  4.5-5.0               0       0.0%            0
+
+──────────────────────────────────────────────────────────────────────────────────────────
+  11. LYRICS vs INSTRUMENTAL POPULARITY
+──────────────────────────────────────────────────────────────────────────────────────────
+
+  Tracks with lyrics:     479,313 (52.8%)
+  Instrumental tracks:    428,928 (47.2%)
+
+──────────────────────────────────────────��───────────────────────────────────────────────
+  12. RECOMMENDED BUCKETING STRATEGIES FOR TRAINING
+──────────────────────────────────────────────────────────────────────────────────────────
+
+  RECOMMENDATION A: Modified User Buckets (merge sparse high-end buckets)
+  ─────────────────────────────────────────────────────────────────
+  The user-proposed buckets are good for the low/mid range but the
+  100K-200K and 200K+ buckets may be very sparse. Consider merging
+  them into a single '100K+' bucket. If each bucket should have ~10K
+  samples, you may also want to split the 0-10 bucket (largest) into
+  sub-ranges or oversample it.
+
+  RECOMMENDATION B: Log-scaled with floor oversample
+  ─────────────────────────────────────────────────────────────────
+  Since popularity follows a power law, use log10-uniform buckets but
+  oversample the 0-play bucket (which represents tracks with no signal).
+  Train with focal loss or weighted sampling to handle imbalance.
+
+  RECOMMENDATION C: Target-transform approach
+  ─────────────────────────────────────────────────────────────────
+  Instead of discrete buckets, train on log(1 + play_count) as a
+  continuous regression target. This naturally compresses the scale.
+  The upvote model can similarly use log(1 + upvote_count).
+  Bucket only for stratified sampling during training, not as labels.
+
+──────────────────────────────────────────────────────────────────────────────────────────
+  13. FEASIBILITY: 10K SAMPLES PER BUCKET (DOWNLOAD ESTIMATE)
+──────────────────────────────────────────────────────────────────────────────────────────
+
+  User-proposed buckets with 10K target per bucket:
+  Bucket              Available   Target   Feasible
+  ────────────────── ────────── ──────── ──────────
+  0-10                  592,886   10,000        YES
+  10-100                207,288   10,000        YES
+  100-1K                 83,443   10,000        YES
+  1K-10K                 20,973   10,000        YES
+  10K-50K                 2,913   10,000 NO (2,913)
+  50K-100K                  331   10,000   NO (331)
+  100K-200K                 225   10,000   NO (225)
+  200K+                     182   10,000   NO (182)
+
+  Total tracks to download: ~43,651
+  Estimated download size:  ~127.9 GB (at ~3MB avg)
+
+==========================================================================================
+  END OF REPORT
+==========================================================================================

public/udio_000032.tar.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d55bc0f54469429277d56c897035b91d18f99b3d0a8ddb4822681e175eb60d04
+size 2679876

search_index/build.log

@@ -0,0 +1,624 @@
+05:06:47 [INFO] ============================================================
+05:06:47 [INFO] ============================================================
+05:06:47 [INFO] LAION-Tunes Search Index Builder
+05:06:47 [INFO] LAION-Tunes Search Index Builder
+05:06:47 [INFO] ============================================================
+05:06:47 [INFO] ============================================================
+05:06:47 [INFO] Found 159 private parquets
+05:06:47 [INFO] Found 159 private parquets
+05:06:47 [INFO] Found 46 output (annotation) parquets
+05:06:47 [INFO] Found 46 output (annotation) parquets
+05:06:47 [INFO] Starting row_id offset: 0
+05:06:47 [INFO] Starting row_id offset: 0
+05:06:47 [INFO]   Created new FAISS tag index
+05:06:47 [INFO]   Created new FAISS tag index
+05:06:47 [INFO]   Created new FAISS lyric index
+05:06:47 [INFO]   Created new FAISS lyric index
+05:06:47 [INFO]   Created new FAISS mood index
+05:06:47 [INFO]   Created new FAISS mood index
+05:06:47 [INFO]   Created new FAISS caption index
+05:06:47 [INFO]   Created new FAISS caption index
+05:06:47 [INFO]   Created new FAISS transcription index
+05:06:47 [INFO]   Created new FAISS transcription index
+05:06:47 [INFO] Processing 159 new private parquets (skipping 0 already done)
+05:06:47 [INFO] Processing 159 new private parquets (skipping 0 already done)
+05:17:50 [INFO] ============================================================
+05:17:50 [INFO] ============================================================
+05:17:50 [INFO] LAION-Tunes Search Index Builder
+05:17:50 [INFO] LAION-Tunes Search Index Builder
+05:17:50 [INFO] ============================================================
+05:17:50 [INFO] ============================================================
+05:17:50 [INFO] Force rebuild requested - starting fresh
+05:17:50 [INFO] Force rebuild requested - starting fresh
+05:17:50 [INFO] Found 159 private parquets
+05:17:50 [INFO] Found 159 private parquets
+05:17:50 [INFO] Found 47 output (annotation) parquets
+05:17:50 [INFO] Found 47 output (annotation) parquets
+05:17:50 [INFO] Starting row_id offset: 0
+05:17:50 [INFO] Starting row_id offset: 0
+05:17:50 [INFO]   Created new FAISS tag index
+05:17:50 [INFO]   Created new FAISS tag index
+05:17:50 [INFO]   Created new FAISS lyric index
+05:17:50 [INFO]   Created new FAISS lyric index
+05:17:50 [INFO]   Created new FAISS mood index
+05:17:50 [INFO]   Created new FAISS mood index
+05:17:50 [INFO]   Created new FAISS caption index
+05:17:50 [INFO]   Created new FAISS caption index
+05:17:50 [INFO]   Created new FAISS transcription index
+05:17:50 [INFO]   Created new FAISS transcription index
+05:17:50 [INFO] Processing 159 new private parquets (skipping 0 already done)
+05:17:50 [INFO] Processing 159 new private parquets (skipping 0 already done)
+05:18:06 [INFO]   [10/159] mureka_000009.tar.parquet: 6479 rows, total=64,612, 4008 rows/s, ETA: 0:04:00
+05:18:06 [INFO]   [10/159] mureka_000009.tar.parquet: 6479 rows, total=64,612, 4008 rows/s, ETA: 0:04:00
+05:18:23 [INFO]   [20/159] mureka_000019.tar.parquet: 7012 rows, total=131,907, 4022 rows/s, ETA: 0:04:02
+05:18:23 [INFO]   [20/159] mureka_000019.tar.parquet: 7012 rows, total=131,907, 4022 rows/s, ETA: 0:04:02
+05:18:40 [INFO]   [30/159] mureka_000029.tar.parquet: 6836 rows, total=199,935, 4000 rows/s, ETA: 0:03:40
+05:18:40 [INFO]   [30/159] mureka_000029.tar.parquet: 6836 rows, total=199,935, 4000 rows/s, ETA: 0:03:40
+05:19:01 [INFO]   [40/159] mureka_000039.tar.parquet: 11439 rows, total=288,781, 4063 rows/s, ETA: 0:05:35
+05:19:01 [INFO]   [40/159] mureka_000039.tar.parquet: 11439 rows, total=288,781, 4063 rows/s, ETA: 0:05:35
+05:19:38 [INFO]   [50/159] riffusion_000000.tar.parquet: 7478 rows, total=391,094, 3642 rows/s, ETA: 0:03:43
+05:19:38 [INFO]   [50/159] riffusion_000000.tar.parquet: 7478 rows, total=391,094, 3642 rows/s, ETA: 0:03:43
+05:21:35 [INFO]   [60/159] riffusion_000010.tar.parquet: 7461 rows, total=465,899, 2076 rows/s, ETA: 0:05:55
+05:21:35 [INFO]   [60/159] riffusion_000010.tar.parquet: 7461 rows, total=465,899, 2076 rows/s, ETA: 0:05:55
+05:22:19 [INFO]   [70/159] suno_000005.tar.parquet: 4746 rows, total=513,585, 1908 rows/s, ETA: 0:03:41
+05:22:19 [INFO]   [70/159] suno_000005.tar.parquet: 4746 rows, total=513,585, 1908 rows/s, ETA: 0:03:41
+05:22:47 [INFO]   [80/159] suno_000015.tar.parquet: 4829 rows, total=561,325, 1891 rows/s, ETA: 0:03:21
+05:22:47 [INFO]   [80/159] suno_000015.tar.parquet: 4829 rows, total=561,325, 1891 rows/s, ETA: 0:03:21
+05:23:17 [INFO]   [90/159] suno_000025.tar.parquet: 4743 rows, total=608,541, 1862 rows/s, ETA: 0:02:55
+05:23:17 [INFO]   [90/159] suno_000025.tar.parquet: 4743 rows, total=608,541, 1862 rows/s, ETA: 0:02:55
+05:23:50 [INFO]   [100/159] suno_000035.tar.parquet: 4719 rows, total=655,939, 1823 rows/s, ETA: 0:02:32
+05:23:50 [INFO]   [100/159] suno_000035.tar.parquet: 4719 rows, total=655,939, 1823 rows/s, ETA: 0:02:32
+05:24:23 [INFO]   [110/159] suno_000045.tar.parquet: 4763 rows, total=704,245, 1792 rows/s, ETA: 0:02:10
+05:24:23 [INFO]   [110/159] suno_000045.tar.parquet: 4763 rows, total=704,245, 1792 rows/s, ETA: 0:02:10
+05:24:40 [INFO]   [120/159] suno_000055.tar.parquet: 4886 rows, total=752,473, 1837 rows/s, ETA: 0:01:43
+05:24:40 [INFO]   [120/159] suno_000055.tar.parquet: 4886 rows, total=752,473, 1837 rows/s, ETA: 0:01:43
+05:24:53 [INFO]   [130/159] udio_000000.tar.parquet: 3387 rows, total=796,488, 1883 rows/s, ETA: 0:00:52
+05:24:53 [INFO]   [130/159] udio_000000.tar.parquet: 3387 rows, total=796,488, 1883 rows/s, ETA: 0:00:52
+05:25:05 [INFO]   [140/159] udio_000012.tar.parquet: 3755 rows, total=833,523, 1916 rows/s, ETA: 0:00:37
+05:25:05 [INFO]   [140/159] udio_000012.tar.parquet: 3755 rows, total=833,523, 1916 rows/s, ETA: 0:00:37
+05:25:19 [INFO]   [150/159] udio_000023.tar.parquet: 4043 rows, total=874,373, 1949 rows/s, ETA: 0:00:18
+05:25:19 [INFO]   [150/159] udio_000023.tar.parquet: 4043 rows, total=874,373, 1949 rows/s, ETA: 0:00:18
+05:25:30 [INFO]   [159/159] udio_000032.tar.parquet: 209 rows, total=908,241, 1975 rows/s, ETA: 0:00:00
+05:25:30 [INFO]   [159/159] udio_000032.tar.parquet: 209 rows, total=908,241, 1975 rows/s, ETA: 0:00:00
+05:25:30 [INFO] 
+Inserted 908,241 new rows (total: 908,241)
+05:25:30 [INFO] 
+Inserted 908,241 new rows (total: 908,241)
+05:25:30 [INFO] Saving FAISS indices...
+05:25:30 [INFO] Saving FAISS indices...
+05:25:32 [INFO]   tag: 908,241 vectors
+05:25:32 [INFO]   tag: 908,241 vectors
+05:25:32 [INFO]   lyric: 479,313 vectors
+05:25:32 [INFO]   lyric: 479,313 vectors
+05:25:33 [INFO]   mood: 383,616 vectors
+05:25:33 [INFO]   mood: 383,616 vectors
+05:25:33 [INFO]   caption: 223,415 vectors
+05:25:33 [INFO]   caption: 223,415 vectors
+05:25:33 [INFO]   transcription: 189,398 vectors
+05:25:33 [INFO]   transcription: 189,398 vectors
+05:25:33 [INFO] Building BM25 indices...
+05:25:33 [INFO] Building BM25 indices...
+05:25:46 [INFO]   tags: 908,241 documents, 164,817 unique tokens
+05:25:46 [INFO]   tags: 908,241 documents, 164,817 unique tokens
+05:26:03 [INFO]   caption: 223,415 documents, 5,755 unique tokens
+05:26:03 [INFO]   caption: 223,415 documents, 5,755 unique tokens
+05:26:28 [INFO]   transcription: 189,398 documents, 875,770 unique tokens
+05:26:28 [INFO]   transcription: 189,398 documents, 875,770 unique tokens
+05:26:28 [INFO]   lyrics_hashed: 0 documents (skipped)
+05:26:28 [INFO]   lyrics_hashed: 0 documents (skipped)
+05:26:28 [INFO] 
+Done in 0:08:38
+05:26:28 [INFO] 
+Done in 0:08:38
+05:26:28 [INFO] Index directory: /home/deployer/laion/music/laion-tunes-final/search_index
+05:26:28 [INFO] Index directory: /home/deployer/laion/music/laion-tunes-final/search_index
+05:26:28 [INFO] Total tracks: 908,241
+05:26:28 [INFO] Total tracks: 908,241
+05:26:28 [INFO]   FAISS tag: 908,241 vectors (2661 MB)
+05:26:28 [INFO]   FAISS tag: 908,241 vectors (2661 MB)
+05:26:28 [INFO]   FAISS lyric: 479,313 vectors (1404 MB)
+05:26:28 [INFO]   FAISS lyric: 479,313 vectors (1404 MB)
+05:26:28 [INFO]   FAISS mood: 383,616 vectors (1124 MB)
+05:26:28 [INFO]   FAISS mood: 383,616 vectors (1124 MB)
+05:26:28 [INFO]   FAISS caption: 223,415 vectors (655 MB)
+05:26:28 [INFO]   FAISS caption: 223,415 vectors (655 MB)
+05:26:28 [INFO]   FAISS transcription: 189,398 vectors (555 MB)
+05:26:28 [INFO]   FAISS transcription: 189,398 vectors (555 MB)
+05:26:28 [INFO]   SQLite: 1327 MB
+05:26:28 [INFO]   SQLite: 1327 MB
+06:39:56 [INFO] Loading search indices...
+06:39:57 [INFO]   FAISS tag: 908,241 vectors
+06:39:58 [INFO]   FAISS lyric: 479,313 vectors
+06:39:58 [INFO]   FAISS mood: 383,616 vectors
+06:39:59 [INFO]   FAISS caption: 223,415 vectors
+06:39:59 [INFO]   FAISS transcription: 189,398 vectors
+06:39:59 [INFO]   BM25 tags: 908,241 documents
+06:39:59 [INFO]   BM25 caption: 223,415 documents
+06:39:59 [INFO]   BM25 transcription: 189,398 documents
+06:39:59 [INFO]   BM25 lyrics_hashed: not found (skipped)
+06:39:59 [INFO]   SQLite: 908,174 tracks
+06:39:59 [INFO] Loading EmbeddingGemma 300M on GPU 0...
+06:40:07 [INFO] Load pretrained SentenceTransformer: google/embeddinggemma-300m
+08:50:24 [INFO] Loading search indices...
+08:50:26 [INFO]   FAISS tag: 908,241 vectors
+08:50:26 [INFO]   FAISS lyric: 479,313 vectors
+08:50:27 [INFO]   FAISS mood: 383,616 vectors
+08:50:27 [INFO]   FAISS caption: 223,415 vectors
+08:50:28 [INFO]   FAISS transcription: 189,398 vectors
+08:50:28 [INFO]   BM25 tags: 908,241 documents
+08:50:28 [INFO]   BM25 caption: 223,415 documents
+08:50:28 [INFO]   BM25 transcription: 189,398 documents
+08:50:28 [INFO]   BM25 lyrics_hashed: not found (skipped)
+08:50:28 [INFO]   SQLite: 908,174 tracks
+08:50:28 [INFO] Loading EmbeddingGemma 300M on GPU 0...
+08:50:36 [INFO] Load pretrained SentenceTransformer: google/embeddinggemma-300m
+08:50:41 [INFO] 14 prompts are loaded, with the keys: ['query', 'document', 'BitextMining', 'Clustering', 'Classification', 'InstructionRetrieval', 'MultilabelClassification', 'PairClassification', 'Reranking', 'Retrieval', 'Retrieval-query', 'Retrieval-document', 'STS', 'Summarization']
+08:50:41 [INFO]   EmbeddingGemma loaded on cuda:0
+09:48:56 [INFO] Loading search indices...
+09:48:58 [INFO]   FAISS tag: 908,241 vectors
+09:48:59 [INFO]   FAISS lyric: 479,313 vectors
+09:49:00 [INFO]   FAISS mood: 383,616 vectors
+09:49:00 [INFO]   FAISS caption: 223,415 vectors
+09:49:00 [INFO]   FAISS transcription: 189,398 vectors
+09:49:00 [INFO]   BM25 tags: 908,241 documents
+09:49:00 [INFO]   BM25 caption: 223,415 documents
+09:49:01 [INFO]   BM25 transcription: 189,398 documents
+09:49:01 [INFO]   BM25 lyrics_hashed: not found (skipped)
+09:49:01 [INFO]   SQLite: 908,174 tracks
+09:49:01 [INFO] Loading EmbeddingGemma 300M on GPU 0...
+09:49:08 [INFO] Load pretrained SentenceTransformer: google/embeddinggemma-300m
+09:49:11 [INFO] 14 prompts are loaded, with the keys: ['query', 'document', 'BitextMining', 'Clustering', 'Classification', 'InstructionRetrieval', 'MultilabelClassification', 'PairClassification', 'Reranking', 'Retrieval', 'Retrieval-query', 'Retrieval-document', 'STS', 'Summarization']
+09:49:11 [INFO]   EmbeddingGemma loaded on cuda:0
+14:39:24 [INFO] ============================================================
+14:39:24 [INFO] Updating search indices with new annotation data
+14:39:24 [INFO] ============================================================
+14:39:24 [INFO] Found 64 output parquets
+14:39:25 [INFO] Current DB: 908,174 tracks, 223,415 captions, 189,398 transcriptions
+14:39:25 [INFO] Building audio_url lookup for tracks needing updates...
+14:39:27 [INFO]   Tracks needing caption: 684,759
+14:39:27 [INFO]   Tracks needing transcription: 718,776
+14:39:32 [INFO]   [10/64] Processed suno_suno_000009.parquet | +0 captions, +0 transcriptions
+14:39:37 [INFO]   [20/64] Processed suno_suno_000019.parquet | +0 captions, +0 transcriptions
+14:39:41 [INFO]   [30/64] Processed suno_suno_000029.parquet | +0 captions, +0 transcriptions
+14:39:46 [INFO]   [40/64] Processed suno_suno_000039.parquet | +0 captions, +0 transcriptions
+14:39:51 [INFO]   [50/64] Processed suno_suno_000049.parquet | +14,381 captions, +12,235 transcriptions
+14:39:57 [INFO]   [60/64] Processed suno_suno_000059.parquet | +62,607 captions, +53,270 transcriptions
+14:39:59 [INFO]   [64/64] Processed suno_suno_000063.parquet | +81,292 captions, +68,409 transcriptions
+14:39:59 [INFO] 
+SQLite updated: +81,292 captions, +68,409 transcriptions
+14:40:00 [INFO]   Captions: 223,415 -> 304,707
+14:40:00 [INFO]   Transcriptions: 189,398 -> 257,807
+14:40:00 [INFO] 
+Rebuilding FAISS caption & transcription indices from all output parquets...
+14:40:06 [INFO]   [10/64] FAISS caption: 47,011, transcription: 40,076
+14:40:11 [INFO]   [20/64] FAISS caption: 94,624, transcription: 80,368
+14:40:17 [INFO]   [30/64] FAISS caption: 141,990, transcription: 120,199
+14:40:23 [INFO]   [40/64] FAISS caption: 190,030, transcription: 160,867
+14:40:29 [INFO]   [50/64] FAISS caption: 237,796, transcription: 201,633
+14:40:35 [INFO]   [60/64] FAISS caption: 286,022, transcription: 242,668
+14:40:37 [INFO]   [64/64] FAISS caption: 304,707, transcription: 257,807
+14:40:37 [INFO] 
+Saving FAISS caption index: 304,707 vectors
+14:40:38 [INFO] Saving FAISS transcription index: 257,807 vectors
+14:40:39 [INFO] 
+Rebuilding all BM25 indices from database...
+14:41:28 [INFO]   BM25 tags: 908,174 docs, 164,817 tokens
+14:41:53 [INFO]   BM25 caption: 304,707 docs, 6,335 tokens
+14:42:24 [INFO]   BM25 transcription: 257,807 docs, 1,142,621 tokens
+14:42:24 [INFO]   BM25 lyrics_hashed: 0 docs (skipped)
+14:42:24 [INFO] 
+Detecting language for new transcriptions...
+14:42:25 [INFO]   162,844 rows need language detection
+14:48:42 [INFO]   Updated language for 162,718 rows
+14:48:42 [INFO] Updating instrumental flags...
+14:48:44 [INFO]   Instrumental tracks: 650,490
+14:48:44 [INFO] 
+Done in 0:09:20
+14:48:44 [INFO] Restart the server to load the updated indices.
+14:50:03 [INFO] Loading search indices...
+14:50:04 [INFO]   FAISS tag: 908,241 vectors
+14:50:05 [INFO]   FAISS lyric: 479,313 vectors
+14:50:06 [INFO]   FAISS mood: 383,616 vectors
+14:50:06 [INFO]   FAISS caption: 304,707 vectors
+14:50:06 [INFO]   FAISS transcription: 257,807 vectors
+14:50:06 [INFO]   BM25 tags: 908,174 documents
+14:50:07 [INFO]   BM25 caption: 304,707 documents
+14:50:07 [INFO]   BM25 transcription: 257,807 documents
+14:50:07 [INFO]   BM25 lyrics_hashed: not found (skipped)
+14:50:07 [INFO]   SQLite: 908,174 tracks
+14:50:07 [INFO] Loading EmbeddingGemma 300M on GPU 0...
+14:50:14 [INFO] Load pretrained SentenceTransformer: google/embeddinggemma-300m
+14:50:17 [INFO] 14 prompts are loaded, with the keys: ['query', 'document', 'BitextMining', 'Clustering', 'Classification', 'InstructionRetrieval', 'MultilabelClassification', 'PairClassification', 'Reranking', 'Retrieval', 'Retrieval-query', 'Retrieval-document', 'STS', 'Summarization']
+14:50:17 [INFO]   EmbeddingGemma loaded on cuda:0
+15:05:52 [INFO] Loading search indices...
+15:05:53 [INFO]   FAISS tag: 908,241 vectors
+15:05:53 [INFO]   FAISS lyric: 479,313 vectors
+15:05:54 [INFO]   FAISS mood: 383,616 vectors
+15:05:54 [INFO]   FAISS caption: 304,707 vectors
+15:05:55 [INFO]   FAISS transcription: 257,807 vectors
+15:05:55 [INFO]   BM25 tags: 908,174 documents
+15:05:55 [INFO]   BM25 caption: 304,707 documents
+15:05:56 [INFO]   BM25 transcription: 257,807 documents
+15:05:56 [INFO]   BM25 lyrics_hashed: not found (skipped)
+15:05:56 [INFO]   SQLite: 908,174 tracks
+15:05:56 [INFO] Loading EmbeddingGemma 300M on GPU 0...
+15:06:03 [INFO] Load pretrained SentenceTransformer: google/embeddinggemma-300m
+15:06:06 [INFO] 14 prompts are loaded, with the keys: ['query', 'document', 'BitextMining', 'Clustering', 'Classification', 'InstructionRetrieval', 'MultilabelClassification', 'PairClassification', 'Reranking', 'Retrieval', 'Retrieval-query', 'Retrieval-document', 'STS', 'Summarization']
+15:06:06 [INFO]   EmbeddingGemma loaded on cuda:0
+15:09:36 [INFO] Loading search indices...
+15:09:37 [INFO]   FAISS tag: 908,241 vectors
+15:09:38 [INFO]   FAISS lyric: 479,313 vectors
+15:09:39 [INFO]   FAISS mood: 383,616 vectors
+15:09:39 [INFO]   FAISS caption: 304,707 vectors
+15:09:40 [INFO]   FAISS transcription: 257,807 vectors
+15:09:40 [INFO]   BM25 tags: 908,174 documents
+15:09:40 [INFO]   BM25 caption: 304,707 documents
+15:09:40 [INFO]   BM25 transcription: 257,807 documents
+15:09:40 [INFO]   BM25 lyrics_hashed: not found (skipped)
+15:09:40 [INFO]   SQLite: 908,174 tracks
+15:09:40 [INFO] Loading EmbeddingGemma 300M on GPU 0...
+15:09:48 [INFO] Load pretrained SentenceTransformer: google/embeddinggemma-300m
+15:09:50 [INFO] 14 prompts are loaded, with the keys: ['query', 'document', 'BitextMining', 'Clustering', 'Classification', 'InstructionRetrieval', 'MultilabelClassification', 'PairClassification', 'Reranking', 'Retrieval', 'Retrieval-query', 'Retrieval-document', 'STS', 'Summarization']
+15:09:50 [INFO]   EmbeddingGemma loaded on cuda:0
+15:17:43 [INFO] Loading search indices...
+15:17:44 [INFO]   FAISS tag: 908,241 vectors
+15:17:45 [INFO]   FAISS lyric: 479,313 vectors
+15:17:45 [INFO]   FAISS mood: 383,616 vectors
+15:17:46 [INFO]   FAISS caption: 304,707 vectors
+15:17:46 [INFO]   FAISS transcription: 257,807 vectors
+15:17:46 [INFO]   BM25 tags: 908,174 documents
+15:17:46 [INFO]   BM25 caption: 304,707 documents
+15:17:47 [INFO]   BM25 transcription: 257,807 documents
+15:17:47 [INFO]   BM25 lyrics_hashed: not found (skipped)
+15:17:47 [INFO]   SQLite: 908,174 tracks
+15:17:47 [INFO] Loading EmbeddingGemma 300M on GPU 0...
+15:17:53 [INFO] Load pretrained SentenceTransformer: google/embeddinggemma-300m
+15:17:56 [INFO] 14 prompts are loaded, with the keys: ['query', 'document', 'BitextMining', 'Clustering', 'Classification', 'InstructionRetrieval', 'MultilabelClassification', 'PairClassification', 'Reranking', 'Retrieval', 'Retrieval-query', 'Retrieval-document', 'STS', 'Summarization']
+15:17:56 [INFO]   EmbeddingGemma loaded on cuda:0
+06:25:39 [INFO] ============================================================
+06:25:39 [INFO] Updating search indices with new annotation data
+06:25:39 [INFO] ============================================================
+06:25:39 [INFO] Found 100 output parquets
+06:25:40 [INFO] Current DB: 908,174 tracks, 304,707 captions, 257,807 transcriptions
+06:25:40 [INFO] Building audio_url lookup for tracks needing updates...
+06:25:42 [INFO]   Tracks needing caption: 603,467
+06:25:42 [INFO]   Tracks needing transcription: 650,367
+06:25:47 [INFO]   [10/100] Processed suno_suno_000007.parquet | +12,846 captions, +11,851 transcriptions
+06:25:52 [INFO]   [20/100] Processed suno_suno_000017.parquet | +12,846 captions, +11,851 transcriptions
+06:25:56 [INFO]   [30/100] Processed suno_suno_000027.parquet | +12,846 captions, +11,851 transcriptions
+06:26:01 [INFO]   [40/100] Processed suno_suno_000037.parquet | +12,846 captions, +11,851 transcriptions
+06:26:06 [INFO]   [50/100] Processed suno_suno_000047.parquet | +12,846 captions, +11,851 transcriptions
+06:26:11 [INFO]   [60/100] Processed suno_suno_000057.parquet | +12,846 captions, +11,851 transcriptions
+06:26:15 [INFO]   [70/100] Processed udio_udio_000002.parquet | +26,016 captions, +22,523 transcriptions
+06:26:20 [INFO]   [80/100] Processed udio_udio_000012.parquet | +56,100 captions, +47,203 transcriptions
+06:26:25 [INFO]   [90/100] Processed udio_udio_000022.parquet | +92,907 captions, +77,389 transcriptions
+06:26:29 [INFO]   [100/100] Processed udio_udio_000032.parquet | +130,818 captions, +108,278 transcriptions
+06:26:29 [INFO] 
+SQLite updated: +130,818 captions, +108,278 transcriptions
+06:26:30 [INFO]   Captions: 304,707 -> 435,525
+06:26:30 [INFO]   Transcriptions: 257,807 -> 366,085
+06:26:30 [INFO] 
+Rebuilding FAISS caption & transcription indices from all output parquets...
+06:26:36 [INFO]   [10/100] FAISS caption: 50,389, transcription: 43,884
+06:26:42 [INFO]   [20/100] FAISS caption: 98,015, transcription: 84,404
+06:26:48 [INFO]   [30/100] FAISS caption: 145,350, transcription: 124,118
+06:26:54 [INFO]   [40/100] FAISS caption: 193,048, transcription: 164,310
+06:27:00 [INFO]   [50/100] FAISS caption: 241,051, transcription: 205,349
+06:27:05 [INFO]   [60/100] FAISS caption: 289,506, transcription: 246,710
+06:27:11 [INFO]   [70/100] FAISS caption: 330,723, transcription: 280,330
+06:27:14 [INFO]   [80/100] FAISS caption: 360,807, transcription: 305,010
+06:27:19 [INFO]   [90/100] FAISS caption: 397,614, transcription: 335,196
+06:27:23 [INFO]   [100/100] FAISS caption: 435,525, transcription: 366,085
+06:27:23 [INFO] 
+Saving FAISS caption index: 435,525 vectors
+06:27:25 [INFO] Saving FAISS transcription index: 366,085 vectors
+06:27:26 [INFO] 
+Rebuilding all BM25 indices from database...
+06:28:24 [INFO]   BM25 tags: 908,174 docs, 164,817 tokens
+06:28:55 [INFO]   BM25 caption: 435,525 docs, 7,222 tokens
+06:29:34 [INFO]   BM25 transcription: 366,085 docs, 1,271,671 tokens
+06:29:34 [INFO]   BM25 lyrics_hashed: 0 docs (skipped)
+06:29:34 [INFO] 
+Detecting language for new transcriptions...
+06:29:34 [INFO]   108,404 rows need language detection
+06:32:51 [INFO]   Updated language for 108,176 rows
+06:32:51 [INFO] Updating instrumental flags...
+06:32:54 [INFO]   Instrumental tracks: 542,314
+06:32:54 [INFO] 
+Done in 0:07:14
+06:32:54 [INFO] Restart the server to load the updated indices.
+06:34:26 [INFO] Loading search indices...
+06:34:27 [INFO]   FAISS tag: 908,241 vectors
+06:34:28 [INFO]   FAISS lyric: 479,313 vectors
+06:34:29 [INFO]   FAISS mood: 383,616 vectors
+06:34:29 [INFO]   FAISS caption: 435,525 vectors
+06:34:30 [INFO]   FAISS transcription: 366,085 vectors
+06:34:30 [INFO]   BM25 tags: 908,174 documents
+06:34:30 [INFO]   BM25 caption: 435,525 documents
+06:34:30 [INFO]   BM25 transcription: 366,085 documents
+06:34:30 [INFO]   BM25 lyrics_hashed: not found (skipped)
+06:34:30 [INFO]   SQLite: 908,174 tracks
+06:34:30 [INFO] Loading EmbeddingGemma 300M on GPU 0...
+06:34:38 [INFO] Load pretrained SentenceTransformer: google/embeddinggemma-300m
+06:34:41 [INFO] 14 prompts are loaded, with the keys: ['query', 'document', 'BitextMining', 'Clustering', 'Classification', 'InstructionRetrieval', 'MultilabelClassification', 'PairClassification', 'Reranking', 'Retrieval', 'Retrieval-query', 'Retrieval-document', 'STS', 'Summarization']
+06:34:41 [INFO]   EmbeddingGemma loaded on cuda:0
+08:07:57 [INFO] ============================================================
+08:07:57 [INFO] Updating search indices with new annotation data
+08:07:57 [INFO] ============================================================
+08:07:57 [INFO] Found 163 output parquets
+08:07:58 [INFO] Current DB: 908,174 tracks, 435,525 captions, 366,085 transcriptions
+08:07:58 [INFO] Building audio_url lookup for tracks needing updates...
+08:08:00 [INFO]   Tracks needing caption: 472,649
+08:08:00 [INFO]   Tracks needing transcription: 542,089
+08:08:06 [INFO]   [10/163] Processed mureka_mureka_000009.parquet | +51,768 captions, +47,726 transcriptions
+08:08:12 [INFO]   [20/163] Processed mureka_mureka_000019.parquet | +109,879 captions, +81,592 transcriptions
+08:08:17 [INFO]   [30/163] Processed mureka_mureka_000029.parquet | +162,145 captions, +96,130 transcriptions
+08:08:25 [INFO]   [40/163] Processed mureka_mureka_000039.parquet | +228,070 captions, +115,240 transcriptions
+08:08:35 [INFO]   [50/163] Processed riffusion_riffusion_000000.parquet | +304,623 captions, +136,881 transcriptions
+08:08:43 [INFO]   [60/163] Processed riffusion_riffusion_000010.parquet | +378,116 captions, +145,795 transcriptions
+08:08:49 [INFO]   [70/163] Processed suno_suno_000004.parquet | +397,475 captions, +148,126 transcriptions
+08:08:54 [INFO]   [80/163] Processed suno_suno_000014.parquet | +397,475 captions, +148,126 transcriptions
+08:08:59 [INFO]   [90/163] Processed suno_suno_000024.parquet | +397,475 captions, +148,126 transcriptions
+08:09:04 [INFO]   [100/163] Processed suno_suno_000034.parquet | +397,475 captions, +148,126 transcriptions
+08:09:09 [INFO]   [110/163] Processed suno_suno_000044.parquet | +397,475 captions, +148,126 transcriptions
+08:09:14 [INFO]   [120/163] Processed suno_suno_000054.parquet | +397,475 captions, +148,126 transcriptions
+08:09:17 [INFO]   [130/163] Processed suno_suno_000064.parquet | +397,475 captions, +148,126 transcriptions
+08:09:20 [INFO]   [140/163] Processed udio_udio_000009.parquet | +397,475 captions, +148,126 transcriptions
+08:09:23 [INFO]   [150/163] Processed udio_udio_000019.parquet | +397,475 captions, +148,126 transcriptions
+08:09:26 [INFO]   [160/163] Processed udio_udio_000029.parquet | +397,475 captions, +148,126 transcriptions
+08:09:27 [INFO]   [163/163] Processed udio_udio_000032.parquet | +397,475 captions, +148,126 transcriptions
+08:09:27 [INFO] 
+SQLite updated: +397,475 captions, +148,126 transcriptions
+08:09:28 [INFO]   Captions: 435,525 -> 832,944
+08:09:28 [INFO]   Transcriptions: 366,085 -> 514,203
+08:09:28 [INFO] 
+Rebuilding FAISS caption & transcription indices from all output parquets...
+08:09:34 [INFO]   [10/163] FAISS caption: 64,614, transcription: 59,577
+08:09:39 [INFO]   [20/163] FAISS caption: 119,762, transcription: 93,119
+08:09:44 [INFO]   [30/163] FAISS caption: 167,165, transcription: 107,127
+08:09:49 [INFO]   [40/163] FAISS caption: 224,029, transcription: 125,404
+08:09:57 [INFO]   [50/163] FAISS caption: 289,465, transcription: 146,141
+08:10:04 [INFO]   [60/163] FAISS caption: 358,160, transcription: 155,047
+08:10:10 [INFO]   [70/163] FAISS caption: 399,456, transcription: 177,220
+08:10:16 [INFO]   [80/163] FAISS caption: 447,113, transcription: 217,934
+08:10:22 [INFO]   [90/163] FAISS caption: 494,415, transcription: 257,780
+08:10:28 [INFO]   [100/163] FAISS caption: 541,837, transcription: 297,257
+08:10:34 [INFO]   [110/163] FAISS caption: 590,099, transcription: 338,537
+08:10:40 [INFO]   [120/163] FAISS caption: 638,204, transcription: 379,841
+08:10:46 [INFO]   [130/163] FAISS caption: 683,718, transcription: 417,468
+08:10:51 [INFO]   [140/163] FAISS caption: 716,565, transcription: 444,277
+08:10:54 [INFO] ============================================================
+08:10:54 [INFO] Updating search indices with new annotation data
+08:10:54 [INFO] ============================================================
+08:10:54 [INFO] Found 163 output parquets
+08:10:55 [INFO] Current DB: 908,174 tracks, 832,944 captions, 514,203 transcriptions
+08:10:55 [INFO] Building audio_url lookup for tracks needing updates...
+08:10:55 [INFO]   [150/163] FAISS caption: 748,649, transcription: 470,637
+08:10:57 [INFO]   Tracks needing caption: 75,230
+08:10:57 [INFO]   Tracks needing transcription: 393,971
+08:11:00 [INFO]   [160/163] FAISS caption: 790,803, transcription: 505,267
+08:11:01 [INFO]   [163/163] FAISS caption: 798,858, transcription: 511,610
+08:11:01 [INFO] 
+Saving FAISS caption index: 798,858 vectors
+08:11:03 [INFO]   [10/163] Processed mureka_mureka_000009.parquet | +0 captions, +0 transcriptions
+08:11:03 [INFO] Saving FAISS transcription index: 511,610 vectors
+08:11:05 [INFO] 
+Rebuilding all BM25 indices from database...
+08:11:09 [INFO]   [20/163] Processed mureka_mureka_000019.parquet | +0 captions, +0 transcriptions
+08:11:14 [INFO]   [30/163] Processed mureka_mureka_000029.parquet | +0 captions, +0 transcriptions
+08:11:20 [INFO]   [40/163] Processed mureka_mureka_000039.parquet | +0 captions, +0 transcriptions
+08:11:26 [INFO]   [50/163] Processed riffusion_riffusion_000000.parquet | +0 captions, +0 transcriptions
+08:11:30 [INFO]   [60/163] Processed riffusion_riffusion_000010.parquet | +0 captions, +0 transcriptions
+08:11:35 [INFO]   [70/163] Processed suno_suno_000004.parquet | +0 captions, +0 transcriptions
+08:11:39 [INFO]   [80/163] Processed suno_suno_000014.parquet | +0 captions, +0 transcriptions
+08:11:44 [INFO]   [90/163] Processed suno_suno_000024.parquet | +0 captions, +0 transcriptions
+08:11:48 [INFO]   [100/163] Processed suno_suno_000034.parquet | +0 captions, +0 transcriptions
+08:11:53 [INFO]   [110/163] Processed suno_suno_000044.parquet | +0 captions, +0 transcriptions
+08:11:58 [INFO]   [120/163] Processed suno_suno_000054.parquet | +0 captions, +0 transcriptions
+08:12:02 [INFO]   [130/163] Processed suno_suno_000064.parquet | +0 captions, +0 transcriptions
+08:12:05 [INFO]   [140/163] Processed udio_udio_000009.parquet | +0 captions, +0 transcriptions
+08:12:09 [INFO]   [150/163] Processed udio_udio_000019.parquet | +0 captions, +0 transcriptions
+08:12:13 [INFO]   [160/163] Processed udio_udio_000029.parquet | +0 captions, +0 transcriptions
+08:12:13 [INFO]   [163/163] Processed udio_udio_000032.parquet | +0 captions, +0 transcriptions
+08:12:13 [INFO] 
+SQLite updated: +0 captions, +0 transcriptions
+08:12:15 [INFO]   Captions: 832,944 -> 832,944
+08:12:15 [INFO]   Transcriptions: 514,203 -> 514,203
+08:12:15 [INFO] 
+Rebuilding FAISS caption & transcription indices from all output parquets...
+08:12:22 [INFO]   [10/163] FAISS caption: 64,614, transcription: 59,577
+08:12:23 [INFO]   BM25 tags: 908,174 docs, 164,817 tokens
+08:12:29 [INFO]   [20/163] FAISS caption: 119,762, transcription: 93,119
+08:12:34 [INFO]   [30/163] FAISS caption: 167,165, transcription: 107,127
+08:12:41 [INFO]   [40/163] FAISS caption: 224,029, transcription: 125,404
+08:12:49 [INFO]   [50/163] FAISS caption: 289,465, transcription: 146,141
+08:12:56 [INFO]   [60/163] FAISS caption: 358,160, transcription: 155,047
+08:13:01 [INFO]   [70/163] FAISS caption: 399,456, transcription: 177,220
+08:13:07 [INFO]   [80/163] FAISS caption: 447,113, transcription: 217,934
+08:13:12 [INFO]   [90/163] FAISS caption: 494,415, transcription: 257,780
+08:13:17 [INFO]   [100/163] FAISS caption: 541,837, transcription: 297,257
+08:13:18 [INFO]   BM25 caption: 832,944 docs, 8,708 tokens
+08:13:23 [INFO]   [110/163] FAISS caption: 590,099, transcription: 338,537
+08:13:29 [INFO]   [120/163] FAISS caption: 638,204, transcription: 379,841
+08:13:34 [INFO]   [130/163] FAISS caption: 683,718, transcription: 417,468
+08:13:39 [INFO]   [140/163] FAISS caption: 716,565, transcription: 444,277
+08:13:43 [INFO]   [150/163] FAISS caption: 748,649, transcription: 470,637
+08:13:47 [INFO]   [160/163] FAISS caption: 790,803, transcription: 505,267
+08:13:48 [INFO]   [163/163] FAISS caption: 798,858, transcription: 511,610
+08:13:48 [INFO] 
+Saving FAISS caption index: 798,858 vectors
+08:13:51 [INFO] Saving FAISS transcription index: 511,610 vectors
+08:13:53 [INFO] 
+Rebuilding all BM25 indices from database...
+08:14:08 [INFO]   BM25 transcription: 514,203 docs, 1,518,385 tokens
+08:14:08 [INFO]   BM25 lyrics_hashed: 0 docs (skipped)
+08:14:08 [INFO] 
+Detecting language for new transcriptions...
+08:14:09 [INFO]   148,346 rows need language detection
+08:15:10 [INFO]   BM25 tags: 908,174 docs, 164,817 tokens
+08:16:04 [INFO]   BM25 caption: 832,944 docs, 8,708 tokens
+08:16:53 [INFO]   BM25 transcription: 514,203 docs, 1,518,385 tokens
+08:16:53 [INFO]   BM25 lyrics_hashed: 0 docs (skipped)
+08:16:53 [INFO] 
+Detecting language for new transcriptions...
+08:16:53 [INFO]   148,346 rows need language detection
+08:18:54 [INFO]   Updated language for 148,068 rows
+08:18:54 [INFO] Updating instrumental flags...
+08:18:57 [INFO]   Instrumental tracks: 394,246
+08:18:57 [INFO] 
+Done in 0:10:59
+08:18:57 [INFO] Restart the server to load the updated indices.
+08:21:37 [INFO]   Updated language for 148,068 rows
+08:21:37 [INFO] Updating instrumental flags...
+08:21:38 [INFO]   Instrumental tracks: 394,246
+08:21:38 [INFO] 
+Done in 0:10:43
+08:21:38 [INFO] Restart the server to load the updated indices.
+08:31:00 [INFO] Loading search indices...
+08:31:02 [INFO]   FAISS tag: 908,241 vectors
+08:31:03 [INFO]   FAISS lyric: 479,313 vectors
+08:31:04 [INFO]   FAISS mood: 383,616 vectors
+08:31:05 [INFO]   FAISS caption: 798,858 vectors
+08:31:06 [INFO]   FAISS transcription: 511,610 vectors
+08:31:06 [INFO]   BM25 tags: 908,174 documents
+08:31:06 [INFO]   BM25 caption: 832,944 documents
+08:31:06 [INFO]   BM25 transcription: 514,203 documents
+08:31:06 [INFO]   BM25 lyrics_hashed: not found (skipped)
+08:31:06 [INFO]   SQLite: 908,174 tracks
+08:31:06 [INFO] Loading EmbeddingGemma 300M on GPU 0...
+08:31:12 [INFO] Load pretrained SentenceTransformer: google/embeddinggemma-300m
+08:31:14 [INFO] 14 prompts are loaded, with the keys: ['query', 'document', 'BitextMining', 'Clustering', 'Classification', 'InstructionRetrieval', 'MultilabelClassification', 'PairClassification', 'Reranking', 'Retrieval', 'Retrieval-query', 'Retrieval-document', 'STS', 'Summarization']
+08:31:14 [INFO]   EmbeddingGemma loaded on cuda:0
+02:57:44 [INFO] Loading search indices...
+02:57:46 [INFO]   FAISS tag: 908,241 vectors
+02:57:46 [INFO]   FAISS lyric: 479,313 vectors
+02:57:47 [INFO]   FAISS mood: 383,616 vectors
+02:57:48 [INFO]   FAISS caption: 798,858 vectors
+02:57:49 [INFO]   FAISS transcription: 511,610 vectors
+02:57:50 [INFO]   FAISS whisper: 465,899 vectors
+02:57:50 [INFO]   Whisper reverse map: 465,832 entries
+02:57:50 [INFO]   BM25 tags: 908,174 documents
+02:57:50 [INFO]   BM25 caption: 832,944 documents
+02:57:51 [INFO]   BM25 transcription: 514,203 documents
+02:57:51 [INFO]   BM25 lyrics_hashed: not found (skipped)
+02:57:51 [INFO]   SQLite: 908,174 tracks
+02:57:51 [INFO] Loading EmbeddingGemma 300M on GPU 0...
+02:57:58 [INFO] Load pretrained SentenceTransformer: google/embeddinggemma-300m
+02:58:00 [INFO] 14 prompts are loaded, with the keys: ['query', 'document', 'BitextMining', 'Clustering', 'Classification', 'InstructionRetrieval', 'MultilabelClassification', 'PairClassification', 'Reranking', 'Retrieval', 'Retrieval-query', 'Retrieval-document', 'STS', 'Summarization']
+02:58:00 [INFO]   EmbeddingGemma loaded on cuda:0
+02:58:00 [INFO] Loading Music-Whisper encoder on GPU 0...
+02:58:01 [INFO]   Music-Whisper encoder loaded on cuda:0 (773 MB)
+03:12:12 [INFO] Loading search indices...
+03:12:13 [INFO]   FAISS tag: 908,241 vectors
+03:12:14 [INFO]   FAISS lyric: 479,313 vectors
+03:12:14 [INFO]   FAISS mood: 383,616 vectors
+03:12:15 [INFO]   FAISS caption: 798,858 vectors
+03:12:16 [INFO]   FAISS transcription: 511,610 vectors
+03:12:17 [INFO]   FAISS whisper: 465,899 vectors
+03:12:17 [INFO]   Whisper reverse map: 465,832 entries
+03:12:17 [INFO]   BM25 tags: 908,174 documents
+03:12:17 [INFO]   BM25 caption: 832,944 documents
+03:12:18 [INFO]   BM25 transcription: 514,203 documents
+03:12:18 [INFO]   BM25 lyrics_hashed: not found (skipped)
+03:12:18 [INFO]   SQLite: 908,174 tracks
+03:12:18 [INFO] Loading EmbeddingGemma 300M on GPU 0...
+03:12:25 [INFO] Load pretrained SentenceTransformer: google/embeddinggemma-300m
+03:12:28 [INFO] 14 prompts are loaded, with the keys: ['query', 'document', 'BitextMining', 'Clustering', 'Classification', 'InstructionRetrieval', 'MultilabelClassification', 'PairClassification', 'Reranking', 'Retrieval', 'Retrieval-query', 'Retrieval-document', 'STS', 'Summarization']
+03:12:28 [INFO]   EmbeddingGemma loaded on cuda:0
+03:12:28 [INFO] Loading Music-Whisper encoder on GPU 0...
+03:12:28 [INFO]   Music-Whisper encoder loaded on cuda:0 (773 MB)
+03:29:51 [INFO] Cached audio embedding: Cupid Stole my Baby.mp3 from 176.4.194.184 (1576ms)
+04:25:41 [INFO] Audio embedding cache hit: Cupid Stole my Baby.mp3 from 176.4.194.184
+05:29:47 [INFO] Cached audio embedding: (Intro).mp3 from 176.4.194.184 (811ms)
+05:33:41 [INFO] Audio embedding cache hit: (Intro).mp3 from 176.4.194.184
+08:31:52 [INFO] Cached audio embedding: (Intro).mp3 from 176.4.194.184 (824ms)
+04:18:22 [INFO] Loading search indices...
+04:18:23 [INFO]   FAISS tag: 908,241 vectors
+04:18:24 [INFO]   FAISS lyric: 479,313 vectors
+04:18:24 [INFO]   FAISS mood: 383,616 vectors
+04:18:25 [INFO]   FAISS caption: 798,858 vectors
+04:18:26 [INFO]   FAISS transcription: 511,610 vectors
+04:18:27 [INFO]   FAISS whisper: 908,174 vectors
+04:18:27 [INFO]   Whisper reverse map: 908,174 entries
+04:18:27 [INFO]   BM25 tags: 908,174 documents
+04:18:28 [INFO]   BM25 caption: 832,944 documents
+04:18:28 [INFO]   BM25 transcription: 514,203 documents
+04:18:28 [INFO]   BM25 lyrics_hashed: not found (skipped)
+04:18:28 [INFO]   SQLite: 908,174 tracks
+04:18:28 [INFO] Loading EmbeddingGemma 300M on GPU 0...
+04:18:33 [INFO] Load pretrained SentenceTransformer: google/embeddinggemma-300m
+04:18:35 [INFO] 14 prompts are loaded, with the keys: ['query', 'document', 'BitextMining', 'Clustering', 'Classification', 'InstructionRetrieval', 'MultilabelClassification', 'PairClassification', 'Reranking', 'Retrieval', 'Retrieval-query', 'Retrieval-document', 'STS', 'Summarization']
+04:18:35 [INFO]   EmbeddingGemma loaded on cuda:0
+04:18:35 [INFO] Loading Music-Whisper encoder on GPU 0...
+04:18:36 [INFO]   Music-Whisper encoder loaded on cuda:0 (773 MB)
+04:49:52 [INFO] Loading search indices...
+04:49:53 [INFO]   FAISS tag: 908,241 vectors
+04:49:54 [INFO]   FAISS lyric: 479,313 vectors
+04:49:55 [INFO]   FAISS mood: 383,616 vectors
+04:49:56 [INFO]   FAISS caption: 798,858 vectors
+04:49:56 [INFO]   FAISS transcription: 511,610 vectors
+04:49:57 [INFO]   FAISS whisper: 908,174 vectors
+04:49:57 [INFO]   Whisper reverse map: 908,174 entries
+04:49:58 [INFO]   BM25 tags: 908,174 documents
+04:49:58 [INFO]   BM25 caption: 832,944 documents
+04:49:58 [INFO]   BM25 transcription: 514,203 documents
+04:49:58 [INFO]   BM25 lyrics_hashed: not found (skipped)
+04:49:58 [INFO]   SQLite: 908,174 tracks
+04:49:58 [INFO] Loading EmbeddingGemma 300M on GPU 0...
+04:50:03 [INFO] Load pretrained SentenceTransformer: google/embeddinggemma-300m
+04:50:05 [INFO] 14 prompts are loaded, with the keys: ['query', 'document', 'BitextMining', 'Clustering', 'Classification', 'InstructionRetrieval', 'MultilabelClassification', 'PairClassification', 'Reranking', 'Retrieval', 'Retrieval-query', 'Retrieval-document', 'STS', 'Summarization']
+04:50:05 [INFO]   EmbeddingGemma loaded on cuda:0
+04:50:05 [INFO] Loading Music-Whisper encoder on GPU 0...
+04:50:06 [INFO]   Music-Whisper encoder loaded on cuda:0 (773 MB)
+04:51:42 [INFO] Loading search indices...
+04:51:43 [INFO]   FAISS tag: 908,241 vectors
+04:51:44 [INFO]   FAISS lyric: 479,313 vectors
+04:51:44 [INFO]   FAISS mood: 383,616 vectors
+04:51:45 [INFO]   FAISS caption: 798,858 vectors
+04:51:46 [INFO]   FAISS transcription: 511,610 vectors
+04:51:47 [INFO]   FAISS whisper: 908,174 vectors
+04:51:47 [INFO]   Whisper reverse map: 908,174 entries
+04:51:47 [INFO]   BM25 tags: 908,174 documents
+04:51:48 [INFO]   BM25 caption: 832,944 documents
+04:51:48 [INFO]   BM25 transcription: 514,203 documents
+04:51:48 [INFO]   BM25 lyrics_hashed: not found (skipped)
+04:51:48 [INFO]   SQLite: 908,174 tracks
+04:51:48 [INFO] Connecting to TEI backend at http://localhost:8090...
+04:51:48 [INFO]   TEI connected at http://localhost:8090 (dim=768)
+04:51:48 [INFO] Loading Music-Whisper encoder on GPU 0...
+04:51:56 [INFO]   Music-Whisper encoder loaded on cuda:0 (182 MB)
+05:02:47 [INFO] Loading search indices...
+05:02:49 [INFO]   FAISS tag: 908,241 vectors
+05:02:49 [INFO]   FAISS lyric: 479,313 vectors
+05:02:50 [INFO]   FAISS mood: 383,616 vectors
+05:02:51 [INFO]   FAISS caption: 798,858 vectors
+05:02:51 [INFO]   FAISS transcription: 511,610 vectors
+05:02:53 [INFO]   FAISS whisper: 908,174 vectors
+05:02:53 [INFO]   Whisper reverse map: 908,174 entries
+05:02:53 [INFO]   BM25 tags: 908,174 documents
+05:02:53 [INFO]   BM25 caption: 832,944 documents
+05:02:53 [INFO]   BM25 transcription: 514,203 documents
+05:02:53 [INFO]   BM25 lyrics_hashed: not found (skipped)
+05:02:53 [INFO]   SQLite: 908,174 tracks
+05:02:53 [INFO] Connecting to TEI backend at http://localhost:8090...
+05:02:54 [INFO]   TEI connected at http://localhost:8090 (dim=768)
+05:02:54 [INFO] Loading Music-Whisper encoder on GPU 0...
+05:02:59 [INFO]   Music-Whisper encoder loaded on cuda:0 (182 MB)
+05:47:06 [INFO] Loading search indices...
+05:47:07 [INFO]   FAISS tag: 908,241 vectors
+05:47:08 [INFO]   FAISS lyric: 479,313 vectors
+05:47:08 [INFO]   FAISS mood: 383,616 vectors
+05:47:09 [INFO]   FAISS caption: 798,858 vectors
+05:47:10 [INFO]   FAISS transcription: 511,610 vectors
+05:47:11 [INFO]   FAISS whisper: 908,174 vectors
+05:47:11 [INFO]   Whisper reverse map: 908,174 entries
+05:47:11 [INFO]   BM25 tags: 908,174 documents
+05:47:11 [INFO]   BM25 caption: 832,944 documents
+05:47:12 [INFO]   BM25 transcription: 514,203 documents
+05:47:12 [INFO]   BM25 lyrics_hashed: not found (skipped)
+05:47:12 [INFO]   SQLite: 908,174 tracks
+05:47:12 [INFO] Connecting to TEI backend at http://localhost:8090...
+05:47:12 [INFO]   TEI connected at http://localhost:8090 (dim=768)
+05:47:12 [INFO] Loading Music-Whisper encoder on GPU 0...
+05:47:18 [INFO]   Music-Whisper encoder loaded on cuda:0 (182 MB)
+07:06:28 [INFO] Cached audio embedding: The Model of My Dreams (Is Closed Source).mp3 from 176.4.189.127 (1631ms)

search_index/build_manifest.json

@@ -0,0 +1,165 @@
+{
+  "processed_private": [
+    "mureka_000000.tar.parquet",
+    "mureka_000001.tar.parquet",
+    "mureka_000002.tar.parquet",
+    "mureka_000003.tar.parquet",
+    "mureka_000004.tar.parquet",
+    "mureka_000005.tar.parquet",
+    "mureka_000006.tar.parquet",
+    "mureka_000007.tar.parquet",
+    "mureka_000008.tar.parquet",
+    "mureka_000009.tar.parquet",
+    "mureka_000010.tar.parquet",
+    "mureka_000011.tar.parquet",
+    "mureka_000012.tar.parquet",
+    "mureka_000013.tar.parquet",
+    "mureka_000014.tar.parquet",
+    "mureka_000015.tar.parquet",
+    "mureka_000016.tar.parquet",
+    "mureka_000017.tar.parquet",
+    "mureka_000018.tar.parquet",
+    "mureka_000019.tar.parquet",
+    "mureka_000020.tar.parquet",
+    "mureka_000021.tar.parquet",
+    "mureka_000022.tar.parquet",
+    "mureka_000023.tar.parquet",
+    "mureka_000024.tar.parquet",
+    "mureka_000025.tar.parquet",
+    "mureka_000026.tar.parquet",
+    "mureka_000027.tar.parquet",
+    "mureka_000028.tar.parquet",
+    "mureka_000029.tar.parquet",
+    "mureka_000030.tar.parquet",
+    "mureka_000031.tar.parquet",
+    "mureka_000032.tar.parquet",
+    "mureka_000033.tar.parquet",
+    "mureka_000034.tar.parquet",
+    "mureka_000035.tar.parquet",
+    "mureka_000036.tar.parquet",
+    "mureka_000037.tar.parquet",
+    "mureka_000038.tar.parquet",
+    "mureka_000039.tar.parquet",
+    "mureka_000040.tar.parquet",
+    "mureka_000041.tar.parquet",
+    "mureka_000042.tar.parquet",
+    "mureka_000043.tar.parquet",
+    "mureka_000044.tar.parquet",
+    "mureka_000045.tar.parquet",
+    "mureka_000046.tar.parquet",
+    "mureka_000047.tar.parquet",
+    "mureka_000048.tar.parquet",
+    "riffusion_000000.tar.parquet",
+    "riffusion_000001.tar.parquet",
+    "riffusion_000002.tar.parquet",
+    "riffusion_000003.tar.parquet",
+    "riffusion_000004.tar.parquet",
+    "riffusion_000005.tar.parquet",
+    "riffusion_000006.tar.parquet",
+    "riffusion_000007.tar.parquet",
+    "riffusion_000008.tar.parquet",
+    "riffusion_000009.tar.parquet",
+    "riffusion_000010.tar.parquet",
+    "riffusion_000011.tar.parquet",
+    "riffusion_000012.tar.parquet",
+    "riffusion_000013.tar.parquet",
+    "sonauto_000001.tar.parquet",
+    "suno_000000.tar.parquet",
+    "suno_000001.tar.parquet",
+    "suno_000002.tar.parquet",
+    "suno_000003.tar.parquet",
+    "suno_000004.tar.parquet",
+    "suno_000005.tar.parquet",
+    "suno_000006.tar.parquet",
+    "suno_000007.tar.parquet",
+    "suno_000008.tar.parquet",
+    "suno_000009.tar.parquet",
+    "suno_000010.tar.parquet",
+    "suno_000011.tar.parquet",
+    "suno_000012.tar.parquet",
+    "suno_000013.tar.parquet",
+    "suno_000014.tar.parquet",
+    "suno_000015.tar.parquet",
+    "suno_000016.tar.parquet",
+    "suno_000017.tar.parquet",
+    "suno_000018.tar.parquet",
+    "suno_000019.tar.parquet",
+    "suno_000020.tar.parquet",
+    "suno_000021.tar.parquet",
+    "suno_000022.tar.parquet",
+    "suno_000023.tar.parquet",
+    "suno_000024.tar.parquet",
+    "suno_000025.tar.parquet",
+    "suno_000026.tar.parquet",
+    "suno_000027.tar.parquet",
+    "suno_000028.tar.parquet",
+    "suno_000029.tar.parquet",
+    "suno_000030.tar.parquet",
+    "suno_000031.tar.parquet",
+    "suno_000032.tar.parquet",
+    "suno_000033.tar.parquet",
+    "suno_000034.tar.parquet",
+    "suno_000035.tar.parquet",
+    "suno_000036.tar.parquet",
+    "suno_000037.tar.parquet",
+    "suno_000038.tar.parquet",
+    "suno_000039.tar.parquet",
+    "suno_000040.tar.parquet",
+    "suno_000041.tar.parquet",
+    "suno_000042.tar.parquet",
+    "suno_000043.tar.parquet",
+    "suno_000044.tar.parquet",
+    "suno_000045.tar.parquet",
+    "suno_000046.tar.parquet",
+    "suno_000047.tar.parquet",
+    "suno_000048.tar.parquet",
+    "suno_000049.tar.parquet",
+    "suno_000050.tar.parquet",
+    "suno_000051.tar.parquet",
+    "suno_000052.tar.parquet",
+    "suno_000053.tar.parquet",
+    "suno_000054.tar.parquet",
+    "suno_000055.tar.parquet",
+    "suno_000056.tar.parquet",
+    "suno_000057.tar.parquet",
+    "suno_000058.tar.parquet",
+    "suno_000059.tar.parquet",
+    "suno_000060.tar.parquet",
+    "suno_000061.tar.parquet",
+    "suno_000062.tar.parquet",
+    "suno_000063.tar.parquet",
+    "suno_000064.tar.parquet",
+    "udio_000000.tar.parquet",
+    "udio_000001.tar.parquet",
+    "udio_000002.tar.parquet",
+    "udio_000003.tar.parquet",
+    "udio_000004.tar.parquet",
+    "udio_000005.tar.parquet",
+    "udio_000006.tar.parquet",
+    "udio_000008.tar.parquet",
+    "udio_000009.tar.parquet",
+    "udio_000010.tar.parquet",
+    "udio_000012.tar.parquet",
+    "udio_000013.tar.parquet",
+    "udio_000014.tar.parquet",
+    "udio_000016.tar.parquet",
+    "udio_000017.tar.parquet",
+    "udio_000018.tar.parquet",
+    "udio_000019.tar.parquet",
+    "udio_000020.tar.parquet",
+    "udio_000021.tar.parquet",
+    "udio_000022.tar.parquet",
+    "udio_000023.tar.parquet",
+    "udio_000024.tar.parquet",
+    "udio_000025.tar.parquet",
+    "udio_000026.tar.parquet",
+    "udio_000027.tar.parquet",
+    "udio_000028.tar.parquet",
+    "udio_000029.tar.parquet",
+    "udio_000030.tar.parquet",
+    "udio_000031.tar.parquet",
+    "udio_000032.tar.parquet"
+  ],
+  "processed_output": [],
+  "total_rows": 908241
+}

search_index/build_stdout.log

@@ -0,0 +1,53 @@
+05:17:50 [INFO] ============================================================
+05:17:50 [INFO] LAION-Tunes Search Index Builder
+05:17:50 [INFO] ============================================================
+05:17:50 [INFO] Force rebuild requested - starting fresh
+05:17:50 [INFO] Found 159 private parquets
+05:17:50 [INFO] Found 47 output (annotation) parquets
+05:17:50 [INFO] Starting row_id offset: 0
+05:17:50 [INFO]   Created new FAISS tag index
+05:17:50 [INFO]   Created new FAISS lyric index
+05:17:50 [INFO]   Created new FAISS mood index
+05:17:50 [INFO]   Created new FAISS caption index
+05:17:50 [INFO]   Created new FAISS transcription index
+05:17:50 [INFO] Processing 159 new private parquets (skipping 0 already done)
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+05:25:30 [INFO] 
+Inserted 908,241 new rows (total: 908,241)
+05:25:30 [INFO] Saving FAISS indices...
+05:25:32 [INFO]   tag: 908,241 vectors
+05:25:32 [INFO]   lyric: 479,313 vectors
+05:25:33 [INFO]   mood: 383,616 vectors
+05:25:33 [INFO]   caption: 223,415 vectors
+05:25:33 [INFO]   transcription: 189,398 vectors
+05:25:33 [INFO] Building BM25 indices...
+05:25:46 [INFO]   tags: 908,241 documents, 164,817 unique tokens
+05:26:03 [INFO]   caption: 223,415 documents, 5,755 unique tokens
+05:26:28 [INFO]   transcription: 189,398 documents, 875,770 unique tokens
+05:26:28 [INFO]   lyrics_hashed: 0 documents (skipped)
+05:26:28 [INFO] 
+Done in 0:08:38
+05:26:28 [INFO] Index directory: /home/deployer/laion/music/laion-tunes-final/search_index
+05:26:28 [INFO] Total tracks: 908,241
+05:26:28 [INFO]   FAISS tag: 908,241 vectors (2661 MB)
+05:26:28 [INFO]   FAISS lyric: 479,313 vectors (1404 MB)
+05:26:28 [INFO]   FAISS mood: 383,616 vectors (1124 MB)
+05:26:28 [INFO]   FAISS caption: 223,415 vectors (655 MB)
+05:26:28 [INFO]   FAISS transcription: 189,398 vectors (555 MB)
+05:26:28 [INFO]   SQLite: 1327 MB

search_index/idmap_lyric.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f35e09d85890824e3f300d8d6e8ad13ebb873d921c541681e10758866e90de3
+size 3834632

search_index/idmap_mood.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61d66c76181687fdca90b106e40d49aa4bb83bc8d48d10a13ee9a921a0c27741
+size 3069056

search_index/idmap_transcription.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f392877dba4341f65a19a2771b638fa531d55a4b65bd803f4af04d0ad5b29830
+size 4093008

server.py

@@ -0,0 +1,1059 @@
+#!/usr/bin/env python3
+"""
+LAION-Tunes Search Server
+==========================
+FastAPI server providing semantic search, BM25 text search, music similarity
+search, and metadata filtering over the LAION-Tunes music dataset.
+
+Features:
+  - Vector similarity search via FAISS (tag/lyric/mood/caption/transcription embeddings)
+  - Music audio similarity search via Whisper encoder mean-pooled embeddings
+  - BM25 text search (tags, caption, transcription, privacy-hashed lyrics)
+  - Combined search: vector search + BM25 + filter by aesthetics/subset + re-rank
+  - Real-time query embedding via EmbeddingGemma 300M
+  - Audio upload → Whisper encoder embedding → FAISS similarity search
+  - Dark-mode HTML frontend with audio players
+
+Usage:
+    python server.py [--port 7860] [--gpu 0] [--host 0.0.0.0]
+
+Requires: search_index/ directory built by build_search_index.py
+"""
+
+import os
+import sys
+import json
+import time
+import hmac
+import hashlib
+import pickle
+import re
+import sqlite3
+import logging
+import argparse
+import tempfile
+import io
+import threading
+from pathlib import Path
+from typing import Optional
+from contextlib import asynccontextmanager
+
+import numpy as np
+import faiss
+from fastapi import FastAPI, Query, UploadFile, File, Form, Request
+from fastapi.responses import HTMLResponse, JSONResponse
+from fastapi.middleware.cors import CORSMiddleware
+from pydantic import BaseModel
+
+# Import BM25Index from the build script (needed for pickle deserialization)
+sys.path.insert(0, str(Path(__file__).parent))
+from build_search_index import BM25Index
+
+# ── Paths ────────────────────────────────────────────────────────────────────
+BASE_DIR = Path(__file__).parent
+INDEX_DIR = BASE_DIR / "search_index"
+HTML_PATH = BASE_DIR / "index.html"
+
+LYRICS_HMAC_KEY = b"laion-tunes-search-2026-secret-key"
+
+log = logging.getLogger("server")
+logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s", datefmt="%H:%M:%S")
+
+# ── Global state (loaded at startup) ─────────────────────────────────────────
+state = {
+    "faiss": {},       # field -> faiss.Index
+    "id_maps": {},     # field -> numpy array of row_ids
+    "bm25": {},        # field -> BM25Index
+    "db_path": None,   # SQLite path
+    "embedder": None,  # SentenceTransformer model
+    "total_tracks": 0,
+    # Whisper encoder for music similarity
+    "whisper_processor": None,
+    "whisper_encoder": None,
+    "whisper_device": None,
+    "whisper_rid_to_idx": {},  # row_id -> faiss_idx for reverse lookup
+    "audio_emb_cache": {},     # (ip, filename) -> {"embedding": np.array, "expires": timestamp}
+}
+
+# ── Audio Embedding Cache ────────────────────────────────────────────────────
+CACHE_TTL = 3600  # 1 hour
+_cache_lock = threading.Lock()
+
+def cache_get(ip, filename):
+    """Get cached audio embedding. Returns numpy array or None."""
+    key = (ip, filename)
+    with _cache_lock:
+        entry = state["audio_emb_cache"].get(key)
+        if entry and entry["expires"] > time.time():
+            return entry["embedding"]
+        # Expired or missing
+        if entry:
+            del state["audio_emb_cache"][key]
+        return None
+
+def cache_set(ip, filename, embedding):
+    """Cache an audio embedding."""
+    key = (ip, filename)
+    with _cache_lock:
+        state["audio_emb_cache"][key] = {
+            "embedding": embedding,
+            "expires": time.time() + CACHE_TTL,
+        }
+        # Evict expired entries (max 1000 entries)
+        if len(state["audio_emb_cache"]) > 1000:
+            now = time.time()
+            expired = [k for k, v in state["audio_emb_cache"].items() if v["expires"] < now]
+            for k in expired:
+                del state["audio_emb_cache"][k]
+
+
+# ── Tokenizer (must match build_search_index.py) ─────────────────────────────
+_TOKEN_RE = re.compile(r"[a-z0-9]+")
+
+def tokenize(text):
+    if not text or not isinstance(text, str):
+        return []
+    return [t for t in _TOKEN_RE.findall(text.lower()) if len(t) >= 2]
+
+def tokenize_and_hash(text, secret_key=LYRICS_HMAC_KEY):
+    tokens = tokenize(text)
+    return [hmac.new(secret_key, t.encode(), hashlib.sha256).hexdigest()[:16] for t in tokens]
+
+
+# ── Startup / Shutdown ───────────────────────────────────────────────────────
+def load_indices():
+    """Load all search indices into memory."""
+    log.info("Loading search indices...")
+
+    # FAISS indices (including whisper)
+    for field in ["tag", "lyric", "mood", "caption", "transcription", "whisper"]:
+        idx_path = INDEX_DIR / f"faiss_{field}.index"
+        map_path = INDEX_DIR / f"idmap_{field}.npy"
+        if idx_path.exists() and map_path.exists():
+            state["faiss"][field] = faiss.read_index(str(idx_path))
+            state["id_maps"][field] = np.load(str(map_path))
+            log.info(f"  FAISS {field}: {state['faiss'][field].ntotal:,} vectors")
+        else:
+            log.info(f"  FAISS {field}: not found (skipped)")
+
+    # Build reverse mapping for whisper (row_id -> faiss_idx)
+    if "whisper" in state["id_maps"]:
+        idmap = state["id_maps"]["whisper"]
+        state["whisper_rid_to_idx"] = {int(rid): idx for idx, rid in enumerate(idmap)}
+        log.info(f"  Whisper reverse map: {len(state['whisper_rid_to_idx']):,} entries")
+
+    # BM25 indices
+    for field in ["tags", "caption", "transcription", "lyrics_hashed"]:
+        pkl_path = INDEX_DIR / f"bm25_{field}.pkl"
+        if pkl_path.exists():
+            with open(pkl_path, "rb") as f:
+                state["bm25"][field] = pickle.load(f)
+            log.info(f"  BM25 {field}: {len(state['bm25'][field].row_ids):,} documents")
+        else:
+            log.info(f"  BM25 {field}: not found (skipped)")
+
+    # SQLite
+    state["db_path"] = str(INDEX_DIR / "metadata.db")
+    conn = sqlite3.connect(state["db_path"])
+    state["total_tracks"] = conn.execute("SELECT COUNT(*) FROM tracks").fetchone()[0]
+    conn.close()
+    log.info(f"  SQLite: {state['total_tracks']:,} tracks")
+
+
+class TEIEmbedder:
+    """Wraps Hugging Face Text Embeddings Inference HTTP API to match SentenceTransformer interface."""
+
+    def __init__(self, url):
+        self.url = url.rstrip("/")
+        import requests as _req
+        self._session = _req.Session()
+        # Warm up / verify
+        r = self._session.post(f"{self.url}/embed", json={"inputs": "test"})
+        r.raise_for_status()
+        self.dim = len(r.json()[0])
+        log.info(f"  TEI connected at {self.url} (dim={self.dim})")
+
+    def encode(self, text, normalize_embeddings=True):
+        r = self._session.post(f"{self.url}/embed", json={"inputs": text})
+        r.raise_for_status()
+        emb = np.array(r.json()[0], dtype=np.float32)
+        if normalize_embeddings:
+            norm = np.linalg.norm(emb)
+            if norm > 0:
+                emb = emb / norm
+        return emb
+
+
+def load_embedder(gpu_id):
+    """Load EmbeddingGemma 300M for real-time query embedding.
+    Supports TEI backend (--embedder tei --tei-url URL) or SentenceTransformer.
+    """
+    tei_url = getattr(app.state, "tei_url", None)
+
+    if tei_url:
+        log.info(f"Connecting to TEI backend at {tei_url}...")
+        state["embedder"] = TEIEmbedder(tei_url)
+        return
+
+    log.info(f"Loading EmbeddingGemma 300M on GPU {gpu_id}...")
+    os.environ["HF_HOME"] = str(Path("/home/deployer/laion/music/.hf_cache"))
+    os.environ["TRANSFORMERS_CACHE"] = str(Path("/home/deployer/laion/music/.hf_cache"))
+
+    import torch
+    from sentence_transformers import SentenceTransformer
+
+    device = f"cuda:{gpu_id}" if torch.cuda.is_available() else "cpu"
+    model = SentenceTransformer(
+        "google/embeddinggemma-300m",
+        device=device,
+        model_kwargs={"torch_dtype": torch.bfloat16},
+    )
+    state["embedder"] = model
+    log.info(f"  EmbeddingGemma loaded on {device}")
+
+
+def load_whisper_encoder(gpu_id):
+    """Load laion/music-whisper encoder for audio similarity search."""
+    if "whisper" not in state["faiss"]:
+        log.info("  Whisper FAISS index not found, skipping encoder load")
+        return
+
+    log.info(f"Loading Music-Whisper encoder on GPU {gpu_id}...")
+    cache_dir = str(Path("/home/deployer/laion/music/.hf_cache_embeddings"))
+    os.environ["HF_HOME"] = cache_dir
+    os.environ["TRANSFORMERS_CACHE"] = cache_dir
+
+    import torch
+    from transformers import WhisperProcessor, WhisperModel
+
+    device = f"cuda:{gpu_id}" if torch.cuda.is_available() else "cpu"
+    processor = WhisperProcessor.from_pretrained("laion/music-whisper", cache_dir=cache_dir)
+    model = WhisperModel.from_pretrained("laion/music-whisper", cache_dir=cache_dir)
+    encoder = model.encoder.to(device).half().eval()
+    del model
+    import torch as _torch
+    _torch.cuda.empty_cache()
+
+    state["whisper_processor"] = processor
+    state["whisper_encoder"] = encoder
+    state["whisper_device"] = device
+    log.info(f"  Music-Whisper encoder loaded on {device} ({_torch.cuda.memory_allocated(gpu_id)/1024**2:.0f} MB)")
+
+    # Restore HF_HOME for EmbeddingGemma
+    os.environ["HF_HOME"] = str(Path("/home/deployer/laion/music/.hf_cache"))
+    os.environ["TRANSFORMERS_CACHE"] = str(Path("/home/deployer/laion/music/.hf_cache"))
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    # Startup
+    load_indices()
+    load_embedder(gpu_id=app.state.gpu_id)
+    if not getattr(app.state, "no_whisper", False):
+        load_whisper_encoder(gpu_id=app.state.gpu_id)
+    else:
+        log.info("Whisper encoder loading skipped (--no-whisper)")
+    yield
+    # Shutdown (nothing to clean up)
+
+
+# ── FastAPI App ──────────────────────────────────────────────────────────────
+app = FastAPI(title="LAION-Tunes Search", lifespan=lifespan)
+app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])
+
+
+# ── Request/Response models ──────────────────────────────────────────────────
+class SearchRequest(BaseModel):
+    query: str
+    negative_query: Optional[str] = None    # Negative prompt for vector search (subtracted from query)
+    search_type: str = "bm25"               # "vector" | "bm25" | "combined"
+    vector_field: str = "caption"           # "tag" | "lyric" | "mood" | "caption" | "transcription"
+    bm25_field: str = "caption"             # "tags" | "caption" | "transcription" | "lyrics_hashed"
+    rank_by: str = "similarity"             # "similarity" | "aesthetics" | "plays" | "likes"
+    min_aesthetics: Optional[float] = None  # Filter: minimum score_average
+    min_similarity: Optional[float] = None  # Filter: minimum cosine similarity
+    subset_filter: Optional[str] = None     # Filter: "suno" | "udio" etc
+    vocal_filter: Optional[str] = None      # "instrumental" | "vocals" | None (all)
+    min_duration: Optional[float] = 60.0    # Minimum duration in seconds (default 1 min)
+    languages: Optional[list[str]] = None   # List of language codes to include, None = all
+    negative_weight: float = 0.7            # Weight for negative query subtraction (0-1)
+    nsfw_filter: Optional[str] = None       # None=all, "sfw_only", "nsfw_only"
+    top_k: int = 50
+    # Two-stage search
+    stage2_enabled: bool = False
+    stage2_query: Optional[str] = None
+    stage2_search_type: str = "vector"      # "vector" | "bm25"
+    stage2_vector_field: str = "caption"
+    stage2_bm25_field: str = "caption"
+    stage2_top_k: int = 50
+
+
+class SimilarSearchRequest(BaseModel):
+    row_id: int
+    top_k: int = 50
+    rank_by: str = "similarity"
+    min_aesthetics: Optional[float] = None
+    subset_filter: Optional[str] = None
+    vocal_filter: Optional[str] = None
+    min_duration: Optional[float] = 60.0
+    languages: Optional[list[str]] = None
+    nsfw_filter: Optional[str] = None
+    # Stage 2
+    stage2_enabled: bool = False
+    stage2_query: Optional[str] = None
+    stage2_search_type: str = "vector"
+    stage2_vector_field: str = "caption"
+    stage2_bm25_field: str = "caption"
+    stage2_top_k: int = 50
+
+
+# ── Helper: fetch metadata from SQLite ───────────────────────────────────────
+def fetch_tracks(row_ids, conn):
+    """Fetch track metadata for given row_ids from SQLite."""
+    if len(row_ids) == 0:
+        return {}
+    placeholders = ",".join("?" * len(row_ids))
+    cursor = conn.execute(
+        f"SELECT * FROM tracks WHERE row_id IN ({placeholders})",
+        list(int(r) for r in row_ids),
+    )
+    columns = [desc[0] for desc in cursor.description]
+    result = {}
+    for row in cursor:
+        d = dict(zip(columns, row))
+        result[d["row_id"]] = d
+    return result
+
+
+def format_result(track, score=None, score_type="similarity", has_whisper_emb=False):
+    """Format a track dict for API response."""
+    # Parse JSON array fields
+    genre = json.loads(track.get("genre_tags") or "[]")
+    scene = json.loads(track.get("scene_tags") or "[]")
+    emotion = json.loads(track.get("emotion_tags") or "[]")
+
+    return {
+        "row_id": track["row_id"],
+        "title": track.get("title") or "Untitled",
+        "audio_url": track.get("audio_url") or "",
+        "subset": track.get("subset") or "",
+        "tags_text": track.get("tags_text") or "",
+        "mood_text": track.get("mood_text") or "",
+        "genre_tags": genre,
+        "scene_tags": scene,
+        "emotion_tags": emotion,
+        "score_average": track.get("score_average"),
+        "score_coherence": track.get("score_coherence"),
+        "score_musicality": track.get("score_musicality"),
+        "score_memorability": track.get("score_memorability"),
+        "score_clarity": track.get("score_clarity"),
+        "score_naturalness": track.get("score_naturalness"),
+        "play_count": track.get("play_count") or 0,
+        "upvote_count": track.get("upvote_count") or 0,
+        "duration_seconds": track.get("duration_seconds"),
+        "music_whisper_caption": track.get("music_whisper_caption") or "",
+        "has_caption": bool(track.get("has_caption")),
+        "has_transcription": bool(track.get("has_transcription")),
+        "is_instrumental": bool(track.get("is_instrumental")),
+        "language": track.get("language") or "unknown",
+        "score": round(float(score), 4) if score is not None else None,
+        "score_type": score_type,
+        "has_whisper_emb": has_whisper_emb,
+        # NSFW safety labels
+        "nsfw_overall_label": track.get("nsfw_overall_label") or "likely_sfw",
+        "nsfw_gore_label": track.get("nsfw_gore_label") or "likely_sfw",
+        "nsfw_sexual_label": track.get("nsfw_sexual_label") or "likely_sfw",
+        "nsfw_hate_label": track.get("nsfw_hate_label") or "likely_sfw",
+        "nsfw_gore_sim": round(float(track["nsfw_gore_sim"]), 4) if track.get("nsfw_gore_sim") is not None else None,
+        "nsfw_sexual_sim": round(float(track["nsfw_sexual_sim"]), 4) if track.get("nsfw_sexual_sim") is not None else None,
+        "nsfw_hate_sim": round(float(track["nsfw_hate_sim"]), 4) if track.get("nsfw_hate_sim") is not None else None,
+    }
+
+
+# ── Search logic ─────────────────────────────────────────────────────────────
+def vector_search(query_embedding, field, top_k):
+    """Search FAISS index, return (row_ids, similarities)."""
+    if field not in state["faiss"]:
+        return np.array([]), np.array([])
+
+    index = state["faiss"][field]
+    id_map = state["id_maps"][field]
+
+    # Reshape for FAISS
+    qvec = query_embedding.reshape(1, -1).astype(np.float32)
+    k = min(top_k, index.ntotal)
+    if k == 0:
+        return np.array([]), np.array([])
+
+    similarities, indices = index.search(qvec, k)
+    similarities = similarities[0]
+    indices = indices[0]
+
+    # Map FAISS indices to row_ids
+    valid = indices >= 0
+    row_ids = id_map[indices[valid]]
+    sims = similarities[valid]
+    return row_ids, sims
+
+
+def bm25_search(query_text, field, top_k):
+    """BM25 text search, return (row_ids, scores)."""
+    if field not in state["bm25"]:
+        return np.array([]), np.array([])
+
+    # Hash tokens for lyrics search
+    if field == "lyrics_hashed":
+        tokens = tokenize_and_hash(query_text)
+    else:
+        tokens = tokenize(query_text)
+
+    return state["bm25"][field].search(tokens, top_k=top_k)
+
+
+def apply_filters(tracks_dict, min_aesthetics=None, subset_filter=None, min_similarity=None,
+                   scores=None, vocal_filter=None, min_duration=None, languages=None,
+                   nsfw_filter=None):
+    """Filter tracks by criteria. Returns filtered row_ids.
+    nsfw_filter: None=all, 'sfw_only'=exclude NSFW, 'nsfw_only'=only NSFW
+    """
+    filtered = []
+    for row_id, track in tracks_dict.items():
+        if min_aesthetics is not None:
+            avg = track.get("score_average")
+            if avg is None or avg < min_aesthetics:
+                continue
+        if subset_filter:
+            if track.get("subset") != subset_filter:
+                continue
+        if min_similarity is not None and scores is not None:
+            sim = scores.get(row_id, 0)
+            if sim < min_similarity:
+                continue
+        if vocal_filter == "instrumental":
+            if not track.get("is_instrumental"):
+                continue
+        elif vocal_filter == "vocals":
+            if track.get("is_instrumental"):
+                continue
+        if min_duration is not None:
+            dur = track.get("duration_seconds")
+            if dur is not None and dur < min_duration:
+                continue
+        if languages:
+            lang = track.get("language") or "unknown"
+            if lang not in languages:
+                continue
+        if nsfw_filter == "sfw_only":
+            label = track.get("nsfw_overall_label") or "likely_sfw"
+            if label != "likely_sfw":
+                continue
+        elif nsfw_filter == "nsfw_only":
+            label = track.get("nsfw_overall_label") or "likely_sfw"
+            if label == "likely_sfw":
+                continue
+        filtered.append(row_id)
+    return filtered
+
+
+def rank_results(row_ids, tracks_dict, rank_by, sim_scores=None):
+    """Re-rank row_ids by the specified field."""
+    if rank_by in ("similarity", "music_similarity") and sim_scores:
+        return sorted(row_ids, key=lambda r: sim_scores.get(r, 0), reverse=True)
+    elif rank_by == "aesthetics":
+        return sorted(row_ids, key=lambda r: tracks_dict.get(r, {}).get("score_average") or 0, reverse=True)
+    elif rank_by == "plays":
+        return sorted(row_ids, key=lambda r: tracks_dict.get(r, {}).get("play_count") or 0, reverse=True)
+    elif rank_by == "likes":
+        return sorted(row_ids, key=lambda r: tracks_dict.get(r, {}).get("upvote_count") or 0, reverse=True)
+    return row_ids
+
+
+# ── Audio embedding helper ──────────────────────────────────────────────────
+def compute_audio_embedding(audio_bytes):
+    """Compute mean-pooled Whisper encoder embedding from audio bytes.
+    Returns L2-normalized 768-dim float32 numpy array.
+    """
+    import torch
+    import librosa
+    import soundfile as sf
+
+    processor = state["whisper_processor"]
+    encoder = state["whisper_encoder"]
+    device = state["whisper_device"]
+
+    if processor is None or encoder is None:
+        raise RuntimeError("Whisper encoder not loaded")
+
+    # Write bytes to temp file, load with librosa
+    with tempfile.NamedTemporaryFile(suffix=".audio", delete=True) as tmp:
+        tmp.write(audio_bytes)
+        tmp.flush()
+        try:
+            audio, sr = librosa.load(tmp.name, sr=16000, mono=True)
+        except Exception:
+            # Try soundfile as fallback
+            audio, sr = sf.read(tmp.name)
+            if len(audio.shape) > 1:
+                audio = audio.mean(axis=1)
+            if sr != 16000:
+                audio = librosa.resample(audio, orig_sr=sr, target_sr=16000)
+
+    # Trim to first 30 seconds
+    max_samples = 30 * 16000
+    audio = audio[:max_samples].astype(np.float32)
+
+    if len(audio) < 1600:  # less than 0.1s
+        raise ValueError("Audio too short (< 0.1 seconds)")
+
+    # Process through Whisper encoder
+    inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
+    input_features = inputs.input_features.to(device).half()
+
+    with torch.no_grad():
+        outputs = encoder(input_features)
+        hidden = outputs.last_hidden_state  # (1, seq, 768)
+        mean_pooled = hidden.mean(dim=1)    # (1, 768)
+        mean_pooled = torch.nn.functional.normalize(mean_pooled, p=2, dim=1)
+        embedding = mean_pooled.cpu().float().numpy()[0]  # (768,)
+
+    return embedding
+
+
+def search_by_whisper_embedding(query_embedding, top_k, rank_by, min_aesthetics,
+                                  subset_filter, vocal_filter, min_duration, languages,
+                                  nsfw_filter=None):
+    """Search whisper FAISS index with a query embedding, apply filters, return formatted results."""
+    t0 = time.time()
+
+    has_filters = (subset_filter or vocal_filter or languages
+                   or (min_aesthetics and min_aesthetics > 0)
+                   or (min_duration and min_duration > 0))
+    fetch_k = max(top_k * 100, 20000) if has_filters else max(top_k * 10, 2000)
+
+    # Search FAISS whisper index
+    row_ids, sims = vector_search(query_embedding, "whisper", fetch_k)
+    sim_scores = {}
+    candidate_row_ids = set()
+    for rid, sim in zip(row_ids, sims):
+        rid = int(rid)
+        sim_scores[rid] = float(sim)
+        candidate_row_ids.add(rid)
+
+    # Fetch metadata
+    conn = sqlite3.connect(state["db_path"])
+    conn.row_factory = sqlite3.Row
+    tracks_dict = fetch_tracks(list(candidate_row_ids), conn)
+
+    # Apply filters
+    filtered_ids = apply_filters(
+        tracks_dict,
+        min_aesthetics=min_aesthetics,
+        subset_filter=subset_filter,
+        vocal_filter=vocal_filter,
+        min_duration=min_duration,
+        languages=languages,
+        nsfw_filter=nsfw_filter,
+    )
+
+    # Rank
+    ranked_ids = rank_results(filtered_ids, tracks_dict, rank_by, sim_scores)
+    final_ids = ranked_ids[:top_k]
+
+    # Build response
+    whisper_rid_set = state["whisper_rid_to_idx"]
+    results = []
+    for rid in final_ids:
+        track = tracks_dict.get(rid)
+        if not track:
+            continue
+        score = sim_scores.get(rid, 0)
+        score_type = "cosine_similarity"
+        if rank_by == "aesthetics":
+            score = track.get("score_average") or 0
+            score_type = "aesthetics"
+        elif rank_by == "plays":
+            score = track.get("play_count") or 0
+            score_type = "play_count"
+        elif rank_by == "likes":
+            score = track.get("upvote_count") or 0
+            score_type = "upvote_count"
+
+        r = format_result(dict(track), score=score, score_type=score_type,
+                          has_whisper_emb=(rid in whisper_rid_set))
+        results.append(r)
+
+    conn.close()
+    search_time_ms = (time.time() - t0) * 1000
+
+    return {
+        "results": results,
+        "total_candidates": len(candidate_row_ids),
+        "total_filtered": len(filtered_ids),
+        "total_tracks": state["total_tracks"],
+        "search_time_ms": round(search_time_ms, 1),
+        "search_type": "music_similarity",
+        "vector_field": "whisper",
+    }
+
+
+# ── API Endpoints ────────────────────────────────────────────────────────────
+@app.get("/", response_class=HTMLResponse)
+async def serve_frontend():
+    """Serve the search frontend."""
+    return HTML_PATH.read_text(encoding="utf-8")
+
+
+@app.get("/nsfw-report", response_class=HTMLResponse)
+async def serve_nsfw_report():
+    """Serve the NSFW safety analysis report."""
+    report = BASE_DIR / "nsfw_safety_report.html"
+    if report.exists():
+        return report.read_text(encoding="utf-8")
+    return HTMLResponse("<h1>Report not generated yet</h1>", status_code=404)
+
+
+@app.post("/api/search")
+async def search(req: SearchRequest):
+    """Main search endpoint supporting vector, BM25, and combined search."""
+    t0 = time.time()
+
+    # Embed query (and optional negative query)
+    t_emb_start = time.time()
+    query_embedding = state["embedder"].encode(
+        req.query, normalize_embeddings=True
+    ).astype(np.float32)
+
+    # Negative prompt: subtract negative embedding from positive, then re-normalize
+    if req.negative_query and req.negative_query.strip() and req.search_type in ("vector", "combined"):
+        neg_embedding = state["embedder"].encode(
+            req.negative_query.strip(), normalize_embeddings=True
+        ).astype(np.float32)
+        weight = max(0.0, min(1.0, req.negative_weight))
+        query_embedding = query_embedding - weight * neg_embedding
+        # Re-normalize to unit length for cosine similarity
+        norm = np.linalg.norm(query_embedding)
+        if norm > 0:
+            query_embedding = query_embedding / norm
+
+    emb_time_ms = (time.time() - t_emb_start) * 1000
+
+    conn = sqlite3.connect(state["db_path"])
+    conn.row_factory = sqlite3.Row
+
+    # Determine candidate pool size - fetch large pool so filters don't starve results.
+    # FAISS IndexFlatIP scans all vectors regardless of k, so large k is ~free.
+    has_filters = (req.subset_filter or req.vocal_filter or req.languages
+                   or (req.min_aesthetics and req.min_aesthetics > 0)
+                   or (req.min_duration and req.min_duration > 0))
+    fetch_k = max(req.top_k * 100, 20000) if has_filters else max(req.top_k * 10, 2000)
+
+    sim_scores = {}  # row_id -> similarity score
+    bm25_scores = {}  # row_id -> bm25 score
+    candidate_row_ids = set()
+
+    if req.search_type in ("vector", "combined"):
+        row_ids, sims = vector_search(query_embedding, req.vector_field, fetch_k)
+        for rid, sim in zip(row_ids, sims):
+            rid = int(rid)
+            sim_scores[rid] = float(sim)
+            candidate_row_ids.add(rid)
+
+    if req.search_type in ("bm25", "combined"):
+        row_ids, scores = bm25_search(req.query, req.bm25_field, fetch_k)
+        for rid, score in zip(row_ids, scores):
+            rid = int(rid)
+            bm25_scores[rid] = float(score)
+            candidate_row_ids.add(rid)
+
+    if req.search_type == "combined" and sim_scores and bm25_scores:
+        # For combined: intersect vector and BM25 candidates
+        candidate_row_ids = set(sim_scores.keys()) & set(bm25_scores.keys())
+        if not candidate_row_ids:
+            # If no intersection, fall back to union
+            candidate_row_ids = set(sim_scores.keys()) | set(bm25_scores.keys())
+
+    # Fetch metadata for candidates
+    tracks_dict = fetch_tracks(list(candidate_row_ids), conn)
+
+    # Use primary scores based on search type for similarity filtering
+    primary_scores = sim_scores if req.search_type in ("vector", "combined") else bm25_scores
+
+    # Apply filters
+    filtered_ids = apply_filters(
+        tracks_dict,
+        min_aesthetics=req.min_aesthetics,
+        subset_filter=req.subset_filter,
+        min_similarity=req.min_similarity,
+        scores=primary_scores,
+        vocal_filter=req.vocal_filter,
+        min_duration=req.min_duration,
+        languages=req.languages,
+        nsfw_filter=req.nsfw_filter,
+    )
+
+    # Rank
+    ranked_ids = rank_results(filtered_ids, tracks_dict, req.rank_by, primary_scores)
+
+    # Take top_k
+    stage1_ids = ranked_ids[:req.top_k]
+
+    # ── Stage 2: re-filter by a second query ────────────────────────────
+    stage2_scores = {}
+    stage2_info = None
+    if req.stage2_enabled and req.stage2_query and req.stage2_query.strip():
+        stage1_set = set(stage1_ids)
+        s2_query = req.stage2_query.strip()
+
+        if req.stage2_search_type == "vector":
+            s2_emb = state["embedder"].encode(s2_query, normalize_embeddings=True).astype(np.float32)
+            # Search with large K to cover stage1 candidates
+            s2_fetch_k = max(len(stage1_set) * 20, 50000)
+            s2_rids, s2_sims = vector_search(s2_emb, req.stage2_vector_field, s2_fetch_k)
+            for rid, sim in zip(s2_rids, s2_sims):
+                rid = int(rid)
+                if rid in stage1_set:
+                    stage2_scores[rid] = float(sim)
+        else:  # bm25
+            s2_fetch_k = max(len(stage1_set) * 20, 50000)
+            s2_rids, s2_scores_arr = bm25_search(s2_query, req.stage2_bm25_field, s2_fetch_k)
+            for rid, sc in zip(s2_rids, s2_scores_arr):
+                rid = int(rid)
+                if rid in stage1_set:
+                    stage2_scores[rid] = float(sc)
+
+        # Rank stage1 candidates by stage2 score, take stage2_top_k
+        stage1_ids = sorted(stage1_ids, key=lambda r: stage2_scores.get(r, -999), reverse=True)
+        stage1_ids = stage1_ids[:req.stage2_top_k]
+        stage2_info = {
+            "query": s2_query,
+            "search_type": req.stage2_search_type,
+            "field": req.stage2_vector_field if req.stage2_search_type == "vector" else req.stage2_bm25_field,
+            "matched": len(stage2_scores),
+            "returned": len(stage1_ids),
+        }
+
+    final_ids = stage1_ids
+
+    # Build response
+    whisper_rid_set = state.get("whisper_rid_to_idx", {})
+    results = []
+    for rid in final_ids:
+        track = tracks_dict.get(rid)
+        if not track:
+            continue
+        # Pick the best score to show
+        if stage2_scores:
+            score = stage2_scores.get(rid, 0)
+            score_type = "cosine_similarity" if req.stage2_search_type == "vector" else "bm25"
+        else:
+            score = primary_scores.get(rid, 0)
+            score_type = "cosine_similarity" if req.search_type == "vector" else "bm25"
+        if req.rank_by == "aesthetics" and not stage2_scores:
+            score = track.get("score_average") or 0
+            score_type = "aesthetics"
+        elif req.rank_by == "plays" and not stage2_scores:
+            score = track.get("play_count") or 0
+            score_type = "play_count"
+        elif req.rank_by == "likes" and not stage2_scores:
+            score = track.get("upvote_count") or 0
+            score_type = "upvote_count"
+
+        r = format_result(dict(track), score=score, score_type=score_type,
+                          has_whisper_emb=(rid in whisper_rid_set))
+        # Attach both scores when stage2 is active
+        if stage2_scores:
+            r["stage1_score"] = round(primary_scores.get(rid, 0), 4)
+            r["stage2_score"] = round(stage2_scores.get(rid, 0), 4)
+        results.append(r)
+
+    conn.close()
+    search_time_ms = (time.time() - t0) * 1000
+
+    resp = {
+        "results": results,
+        "total_candidates": len(candidate_row_ids),
+        "total_filtered": len(filtered_ids),
+        "total_tracks": state["total_tracks"],
+        "search_time_ms": round(search_time_ms, 1),
+        "query_embedding_time_ms": round(emb_time_ms, 1),
+        "search_type": req.search_type,
+        "vector_field": req.vector_field,
+        "bm25_field": req.bm25_field,
+    }
+    if stage2_info:
+        resp["stage2"] = stage2_info
+    return resp
+
+
+@app.post("/api/search_by_audio")
+async def search_by_audio(
+    request: Request,
+    audio: UploadFile = File(...),
+    top_k: int = Form(50),
+    rank_by: str = Form("similarity"),
+    subset_filter: Optional[str] = Form(None),
+    vocal_filter: Optional[str] = Form(None),
+    min_duration: Optional[float] = Form(None),
+    min_aesthetics: Optional[float] = Form(None),
+    languages: Optional[str] = Form(None),  # comma-separated language codes
+    nsfw_filter: Optional[str] = Form(None),  # None=all, "sfw_only", "nsfw_only"
+    # Stage 2 fields
+    stage2_enabled: Optional[str] = Form(None),
+    stage2_query: Optional[str] = Form(None),
+    stage2_search_type: Optional[str] = Form("vector"),
+    stage2_vector_field: Optional[str] = Form("caption"),
+    stage2_bm25_field: Optional[str] = Form("caption"),
+    stage2_top_k: int = Form(50),
+):
+    """Search by uploaded audio: compute whisper embedding, find nearest neighbors."""
+    if "whisper" not in state["faiss"]:
+        return JSONResponse(status_code=503, content={"error": "Whisper index not available"})
+    if state["whisper_encoder"] is None:
+        return JSONResponse(status_code=503, content={"error": "Whisper encoder not loaded"})
+
+    t0 = time.time()
+
+    # Client IP for caching
+    client_ip = request.client.host if request.client else "unknown"
+    filename = audio.filename or "unknown"
+
+    # Check cache
+    cached_emb = cache_get(client_ip, filename)
+    cache_hit = cached_emb is not None
+
+    if cache_hit:
+        query_embedding = cached_emb
+        emb_time_ms = 0.0
+        log.info(f"Audio embedding cache hit: {filename} from {client_ip}")
+    else:
+        # Read audio bytes
+        audio_bytes = await audio.read()
+        if len(audio_bytes) == 0:
+            return JSONResponse(status_code=400, content={"error": "Empty audio file"})
+        if len(audio_bytes) > 100 * 1024 * 1024:  # 100 MB limit
+            return JSONResponse(status_code=400, content={"error": "File too large (max 100 MB)"})
+
+        # Compute embedding
+        t_emb = time.time()
+        try:
+            query_embedding = compute_audio_embedding(audio_bytes)
+        except Exception as e:
+            return JSONResponse(status_code=400, content={"error": f"Audio processing failed: {str(e)}"})
+        emb_time_ms = (time.time() - t_emb) * 1000
+
+        # Cache the result
+        cache_set(client_ip, filename, query_embedding)
+        log.info(f"Cached audio embedding: {filename} from {client_ip} ({emb_time_ms:.0f}ms)")
+
+    # Parse languages
+    lang_list = None
+    if languages and languages.strip():
+        lang_list = [l.strip() for l in languages.split(",") if l.strip()]
+
+    # Parse optional floats that come as "None" strings from form
+    min_dur = min_duration if min_duration and min_duration > 0 else None
+    min_aes = min_aesthetics if min_aesthetics and min_aesthetics > 0 else None
+    sub_filter = subset_filter if subset_filter and subset_filter != "null" else None
+    voc_filter = vocal_filter if vocal_filter and vocal_filter != "null" else None
+
+    nsfw_f = nsfw_filter if nsfw_filter and nsfw_filter not in ("null", "") else None
+
+    result = search_by_whisper_embedding(
+        query_embedding, top_k, rank_by, min_aes, sub_filter, voc_filter, min_dur, lang_list,
+        nsfw_filter=nsfw_f,
+    )
+    result["query_embedding_time_ms"] = round(emb_time_ms, 1)
+    result["audio_filename"] = filename
+    result["cache_hit"] = cache_hit
+
+    # Stage 2: re-filter music similarity results by text query
+    s2_enabled = stage2_enabled and stage2_enabled.lower() in ("true", "1", "yes")
+    if s2_enabled and stage2_query and stage2_query.strip():
+        stage1_results = result["results"]
+        stage1_set = {r["row_id"] for r in stage1_results}
+        s2_query = stage2_query.strip()
+        stage2_scores = {}
+
+        if stage2_search_type == "vector":
+            s2_emb = state["embedder"].encode(s2_query, normalize_embeddings=True).astype(np.float32)
+            s2_fetch_k = max(len(stage1_set) * 20, 50000)
+            s2_rids, s2_sims = vector_search(s2_emb, stage2_vector_field, s2_fetch_k)
+            for rid, sim in zip(s2_rids, s2_sims):
+                rid = int(rid)
+                if rid in stage1_set:
+                    stage2_scores[rid] = float(sim)
+        else:  # bm25
+            s2_fetch_k = max(len(stage1_set) * 20, 50000)
+            s2_rids, s2_scores_arr = bm25_search(s2_query, stage2_bm25_field, s2_fetch_k)
+            for rid, sc in zip(s2_rids, s2_scores_arr):
+                rid = int(rid)
+                if rid in stage1_set:
+                    stage2_scores[rid] = float(sc)
+
+        # Re-rank by stage2 score
+        for r in stage1_results:
+            r["stage1_score"] = r.get("score", 0)
+            r["stage2_score"] = round(stage2_scores.get(r["row_id"], -999), 4)
+
+        stage1_results.sort(key=lambda r: r["stage2_score"], reverse=True)
+        result["results"] = stage1_results[:stage2_top_k]
+        result["stage2"] = {
+            "query": s2_query,
+            "search_type": stage2_search_type,
+            "field": stage2_vector_field if stage2_search_type == "vector" else stage2_bm25_field,
+            "matched": len(stage2_scores),
+            "returned": len(result["results"]),
+        }
+
+    return result
+
+
+@app.post("/api/search_similar")
+async def search_similar(req: SimilarSearchRequest):
+    """Search for tracks similar to a given sample using pre-computed whisper embeddings."""
+    if "whisper" not in state["faiss"]:
+        return JSONResponse(status_code=503, content={"error": "Whisper index not available"})
+
+    rid_to_idx = state["whisper_rid_to_idx"]
+    if req.row_id not in rid_to_idx:
+        return JSONResponse(status_code=404, content={
+            "error": f"No whisper embedding found for row_id {req.row_id}. "
+                     f"Only {len(rid_to_idx):,} of {state['total_tracks']:,} tracks have whisper embeddings."
+        })
+
+    t0 = time.time()
+
+    # Reconstruct the embedding from FAISS index
+    faiss_idx = rid_to_idx[req.row_id]
+    query_embedding = state["faiss"]["whisper"].reconstruct(faiss_idx)
+
+    result = search_by_whisper_embedding(
+        query_embedding, req.top_k, req.rank_by,
+        req.min_aesthetics, req.subset_filter, req.vocal_filter,
+        req.min_duration, req.languages,
+        nsfw_filter=req.nsfw_filter,
+    )
+    result["reference_row_id"] = req.row_id
+
+    # Look up title for the reference
+    conn = sqlite3.connect(state["db_path"])
+    row = conn.execute("SELECT title FROM tracks WHERE row_id=?", (req.row_id,)).fetchone()
+    conn.close()
+    if row:
+        result["reference_title"] = row[0] or "Untitled"
+
+    # Stage 2: re-filter music similarity results by text query
+    if req.stage2_enabled and req.stage2_query and req.stage2_query.strip():
+        stage1_results = result["results"]
+        stage1_set = {r["row_id"] for r in stage1_results}
+        s2_query = req.stage2_query.strip()
+        stage2_scores = {}
+
+        if req.stage2_search_type == "vector":
+            s2_emb = state["embedder"].encode(s2_query, normalize_embeddings=True).astype(np.float32)
+            s2_fetch_k = max(len(stage1_set) * 20, 50000)
+            s2_rids, s2_sims = vector_search(s2_emb, req.stage2_vector_field, s2_fetch_k)
+            for rid, sim in zip(s2_rids, s2_sims):
+                rid = int(rid)
+                if rid in stage1_set:
+                    stage2_scores[rid] = float(sim)
+        else:  # bm25
+            s2_fetch_k = max(len(stage1_set) * 20, 50000)
+            s2_rids, s2_scores_arr = bm25_search(s2_query, req.stage2_bm25_field, s2_fetch_k)
+            for rid, sc in zip(s2_rids, s2_scores_arr):
+                rid = int(rid)
+                if rid in stage1_set:
+                    stage2_scores[rid] = float(sc)
+
+        for r in stage1_results:
+            r["stage1_score"] = r.get("score", 0)
+            r["stage2_score"] = round(stage2_scores.get(r["row_id"], -999), 4)
+
+        stage1_results.sort(key=lambda r: r["stage2_score"], reverse=True)
+        result["results"] = stage1_results[:req.stage2_top_k]
+        result["stage2"] = {
+            "query": s2_query,
+            "search_type": req.stage2_search_type,
+            "field": req.stage2_vector_field if req.stage2_search_type == "vector" else req.stage2_bm25_field,
+            "matched": len(stage2_scores),
+            "returned": len(result["results"]),
+        }
+
+    return result
+
+
+@app.get("/api/stats")
+async def stats():
+    """Dataset statistics."""
+    conn = sqlite3.connect(state["db_path"])
+    total = conn.execute("SELECT COUNT(*) FROM tracks").fetchone()[0]
+
+    subsets = {}
+    for row in conn.execute("SELECT subset, COUNT(*) FROM tracks GROUP BY subset"):
+        subsets[row[0]] = row[1]
+
+    avg_scores = {}
+    for row in conn.execute(
+        "SELECT AVG(score_average), MIN(score_average), MAX(score_average) FROM tracks WHERE score_average IS NOT NULL"
+    ):
+        avg_scores = {"mean": round(row[0], 3) if row[0] else None, "min": row[1], "max": row[2]}
+
+    with_caption = conn.execute("SELECT COUNT(*) FROM tracks WHERE has_caption=1").fetchone()[0]
+    with_transcription = conn.execute("SELECT COUNT(*) FROM tracks WHERE has_transcription=1").fetchone()[0]
+
+    faiss_stats = {}
+    for field, idx in state["faiss"].items():
+        faiss_stats[field] = idx.ntotal
+
+    bm25_stats = {}
+    for field, idx in state["bm25"].items():
+        bm25_stats[field] = len(idx.row_ids)
+
+    # Language distribution
+    languages = {}
+    for row in conn.execute(
+        "SELECT language, COUNT(*) FROM tracks WHERE language IS NOT NULL AND language != '' "
+        "GROUP BY language ORDER BY COUNT(*) DESC"
+    ):
+        languages[row[0]] = row[1]
+
+    instrumental = conn.execute("SELECT COUNT(*) FROM tracks WHERE is_instrumental=1").fetchone()[0]
+
+    conn.close()
+
+    return {
+        "total_tracks": total,
+        "subsets": subsets,
+        "score_average": avg_scores,
+        "with_caption": with_caption,
+        "with_transcription": with_transcription,
+        "faiss_indices": faiss_stats,
+        "bm25_indices": bm25_stats,
+        "languages": languages,
+        "instrumental_count": instrumental,
+        "whisper_embeddings": state["faiss"]["whisper"].ntotal if "whisper" in state["faiss"] else 0,
+    }
+
+
+# ── Main ─────────────────────────────────────────────────────────────────────
+def main():
+    parser = argparse.ArgumentParser(description="LAION-Tunes Search Server")
+    parser.add_argument("--port", type=int, default=7860)
+    parser.add_argument("--host", type=str, default="0.0.0.0")
+    parser.add_argument("--gpu", type=int, default=0, help="GPU ID for EmbeddingGemma")
+    parser.add_argument("--tei-url", type=str, default=None,
+                        help="Use TEI backend for embeddings (e.g. http://localhost:8090)")
+    parser.add_argument("--no-whisper", action="store_true",
+                        help="Skip loading whisper encoder (for CPU-only deployment)")
+    args = parser.parse_args()
+
+    app.state.gpu_id = args.gpu
+    app.state.tei_url = args.tei_url
+    app.state.no_whisper = args.no_whisper
+
+    import uvicorn
+    uvicorn.run(app, host=args.host, port=args.port, log_level="info")
+
+
+if __name__ == "__main__":
+    main()

update_indices.py

@@ -0,0 +1,285 @@
+#!/usr/bin/env python3
+"""
+Update search indices with new annotation data.
+
+Reads all output parquets, updates SQLite rows that are missing captions/transcriptions,
+rebuilds FAISS caption & transcription indices, and rebuilds all BM25 indices.
+Preserves existing language/is_instrumental columns.
+"""
+import os
+import sys
+import json
+import glob
+import time
+import hmac
+import hashlib
+import pickle
+import re
+import sqlite3
+import logging
+from pathlib import Path
+from datetime import timedelta
+from collections import defaultdict
+
+import numpy as np
+import pandas as pd
+import faiss
+
+sys.path.insert(0, str(Path(__file__).parent))
+from build_search_index import BM25Index, tokenize, tokenize_and_hash
+
+BASE_DIR = Path(__file__).parent
+OUTPUT_DIR = Path("/home/deployer/laion/music/output")
+INDEX_DIR = BASE_DIR / "search_index"
+DB_PATH = INDEX_DIR / "metadata.db"
+LYRICS_HMAC_KEY = b"laion-tunes-search-2026-secret-key"
+
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s [%(levelname)s] %(message)s",
+    datefmt="%H:%M:%S",
+    handlers=[logging.StreamHandler(sys.stdout)],
+)
+log = logging.getLogger("updater")
+
+
+def update_all():
+    t0 = time.time()
+
+    # ── Step 1: Update SQLite with new captions/transcriptions ─────────
+    log.info("=" * 60)
+    log.info("Updating search indices with new annotation data")
+    log.info("=" * 60)
+
+    output_files = sorted(glob.glob(str(OUTPUT_DIR / "*.parquet")))
+    log.info(f"Found {len(output_files)} output parquets")
+
+    conn = sqlite3.connect(str(DB_PATH))
+    conn.execute("PRAGMA journal_mode=WAL")
+    conn.execute("PRAGMA synchronous=NORMAL")
+
+    # Get current counts
+    total_tracks = conn.execute("SELECT COUNT(*) FROM tracks").fetchone()[0]
+    existing_captions = conn.execute("SELECT COUNT(*) FROM tracks WHERE has_caption=1").fetchone()[0]
+    existing_transcriptions = conn.execute("SELECT COUNT(*) FROM tracks WHERE has_transcription=1").fetchone()[0]
+    log.info(f"Current DB: {total_tracks:,} tracks, {existing_captions:,} captions, {existing_transcriptions:,} transcriptions")
+
+    # Build lookup: audio_url -> row_id for tracks missing captions
+    log.info("Building audio_url lookup for tracks needing updates...")
+    url_to_rowid = {}
+    cursor = conn.execute("SELECT row_id, audio_url, has_caption, has_transcription FROM tracks")
+    needs_caption = set()
+    needs_transcription = set()
+    for row_id, audio_url, has_cap, has_trans in cursor:
+        url_to_rowid[audio_url] = row_id
+        if not has_cap:
+            needs_caption.add(audio_url)
+        if not has_trans:
+            needs_transcription.add(audio_url)
+    log.info(f"  Tracks needing caption: {len(needs_caption):,}")
+    log.info(f"  Tracks needing transcription: {len(needs_transcription):,}")
+
+    # Process each output parquet
+    updated_captions = 0
+    updated_transcriptions = 0
+
+    for fi, out_path in enumerate(output_files):
+        out_name = Path(out_path).name
+        try:
+            out_df = pd.read_parquet(out_path)
+        except Exception as e:
+            log.warning(f"  Failed to read {out_name}: {e}")
+            continue
+
+        batch_updates = []
+        for _, row in out_df.iterrows():
+            audio_url = str(row.get("audio_url", ""))
+            if not audio_url or audio_url not in url_to_rowid:
+                continue
+
+            row_id = url_to_rowid[audio_url]
+            caption = str(row.get("music_whisper_caption", "")) if pd.notna(row.get("music_whisper_caption")) else ""
+            transcription = str(row.get("parakeet_transcription", "")) if pd.notna(row.get("parakeet_transcription")) else ""
+
+            updates = {}
+            if caption and audio_url in needs_caption:
+                updates["music_whisper_caption"] = caption
+                updates["has_caption"] = 1
+                updated_captions += 1
+            if transcription and audio_url in needs_transcription:
+                updates["parakeet_transcription"] = transcription
+                updates["has_transcription"] = 1
+                updated_transcriptions += 1
+
+            if updates:
+                set_clause = ", ".join(f"{k}=?" for k in updates.keys())
+                vals = list(updates.values()) + [row_id]
+                conn.execute(f"UPDATE tracks SET {set_clause} WHERE row_id=?", vals)
+
+        conn.commit()
+
+        if (fi + 1) % 10 == 0 or (fi + 1) == len(output_files):
+            log.info(f"  [{fi+1}/{len(output_files)}] Processed {out_name} | "
+                     f"+{updated_captions:,} captions, +{updated_transcriptions:,} transcriptions")
+
+    log.info(f"\nSQLite updated: +{updated_captions:,} captions, +{updated_transcriptions:,} transcriptions")
+
+    # Verify new counts
+    new_captions = conn.execute("SELECT COUNT(*) FROM tracks WHERE has_caption=1").fetchone()[0]
+    new_transcriptions = conn.execute("SELECT COUNT(*) FROM tracks WHERE has_transcription=1").fetchone()[0]
+    log.info(f"  Captions: {existing_captions:,} -> {new_captions:,}")
+    log.info(f"  Transcriptions: {existing_transcriptions:,} -> {new_transcriptions:,}")
+
+    # ── Step 2: Rebuild FAISS caption & transcription indices ──────────
+    log.info("\nRebuilding FAISS caption & transcription indices from all output parquets...")
+
+    faiss_caption = faiss.IndexFlatIP(768)
+    faiss_transcription = faiss.IndexFlatIP(768)
+    idmap_caption = []
+    idmap_transcription = []
+
+    for fi, out_path in enumerate(output_files):
+        out_name = Path(out_path).name
+        try:
+            out_df = pd.read_parquet(out_path)
+        except Exception:
+            continue
+
+        cap_vecs, cap_ids = [], []
+        trans_vecs, trans_ids = [], []
+
+        for _, row in out_df.iterrows():
+            audio_url = str(row.get("audio_url", ""))
+            if audio_url not in url_to_rowid:
+                continue
+            row_id = url_to_rowid[audio_url]
+
+            # Caption embedding
+            emb = row.get("caption_embedding")
+            if emb is not None and isinstance(emb, np.ndarray) and len(emb) == 768:
+                vec = emb.astype(np.float32)
+                norm = np.linalg.norm(vec)
+                if norm > 0:
+                    vec /= norm
+                    cap_vecs.append(vec)
+                    cap_ids.append(row_id)
+
+            # Transcription embedding
+            emb = row.get("transcription_embedding")
+            if emb is not None and isinstance(emb, np.ndarray) and len(emb) == 768:
+                vec = emb.astype(np.float32)
+                norm = np.linalg.norm(vec)
+                if norm > 0:
+                    vec /= norm
+                    trans_vecs.append(vec)
+                    trans_ids.append(row_id)
+
+        if cap_vecs:
+            faiss_caption.add(np.stack(cap_vecs))
+            idmap_caption.extend(cap_ids)
+        if trans_vecs:
+            faiss_transcription.add(np.stack(trans_vecs))
+            idmap_transcription.extend(trans_ids)
+
+        if (fi + 1) % 10 == 0 or (fi + 1) == len(output_files):
+            log.info(f"  [{fi+1}/{len(output_files)}] FAISS caption: {faiss_caption.ntotal:,}, "
+                     f"transcription: {faiss_transcription.ntotal:,}")
+
+    # Save new FAISS indices
+    log.info(f"\nSaving FAISS caption index: {faiss_caption.ntotal:,} vectors")
+    faiss.write_index(faiss_caption, str(INDEX_DIR / "faiss_caption.index"))
+    np.save(str(INDEX_DIR / "idmap_caption.npy"), np.array(idmap_caption, dtype=np.int64))
+
+    log.info(f"Saving FAISS transcription index: {faiss_transcription.ntotal:,} vectors")
+    faiss.write_index(faiss_transcription, str(INDEX_DIR / "faiss_transcription.index"))
+    np.save(str(INDEX_DIR / "idmap_transcription.npy"), np.array(idmap_transcription, dtype=np.int64))
+
+    # ── Step 3: Rebuild all BM25 indices from DB ──────────────────────
+    log.info("\nRebuilding all BM25 indices from database...")
+
+    bm25_data = {
+        "tags": {"docs": [], "row_ids": []},
+        "caption": {"docs": [], "row_ids": []},
+        "transcription": {"docs": [], "row_ids": []},
+        "lyrics_hashed": {"docs": [], "row_ids": []},
+    }
+
+    cursor = conn.execute(
+        "SELECT row_id, tags_text, music_whisper_caption, parakeet_transcription FROM tracks"
+    )
+    for row_id, tags, caption, transcript in cursor:
+        if tags:
+            bm25_data["tags"]["docs"].append(tokenize(tags))
+            bm25_data["tags"]["row_ids"].append(row_id)
+        if caption:
+            bm25_data["caption"]["docs"].append(tokenize(caption))
+            bm25_data["caption"]["row_ids"].append(row_id)
+        if transcript:
+            bm25_data["transcription"]["docs"].append(tokenize(transcript))
+            bm25_data["transcription"]["row_ids"].append(row_id)
+
+    for bm25_name, bdata in bm25_data.items():
+        if bdata["docs"]:
+            idx = BM25Index()
+            idx.build(bdata["docs"], bdata["row_ids"])
+            with open(INDEX_DIR / f"bm25_{bm25_name}.pkl", "wb") as f:
+                pickle.dump(idx, f, protocol=pickle.HIGHEST_PROTOCOL)
+            log.info(f"  BM25 {bm25_name}: {len(bdata['docs']):,} docs, {len(idx.vocab):,} tokens")
+        else:
+            log.info(f"  BM25 {bm25_name}: 0 docs (skipped)")
+
+    # ── Step 4: Update language detection for new transcriptions ──────
+    log.info("\nDetecting language for new transcriptions...")
+    try:
+        from langdetect import detect, LangDetectException
+
+        new_lang_rows = conn.execute(
+            "SELECT row_id, parakeet_transcription FROM tracks "
+            "WHERE has_transcription=1 AND (language IS NULL OR language='' OR language='unknown')"
+        ).fetchall()
+        log.info(f"  {len(new_lang_rows):,} rows need language detection")
+
+        updates = []
+        for row_id, text in new_lang_rows:
+            if not text or len(text.strip()) < 10:
+                continue
+            try:
+                lang = detect(text[:300])
+                updates.append((lang, row_id))
+            except LangDetectException:
+                pass
+
+        if updates:
+            conn.executemany("UPDATE tracks SET language=? WHERE row_id=?", updates)
+            conn.commit()
+            log.info(f"  Updated language for {len(updates):,} rows")
+    except ImportError:
+        log.warning("  langdetect not installed, skipping language detection")
+
+    # ── Step 5: Update instrumental flags ─────────────────────────────
+    log.info("Updating instrumental flags...")
+    conn.execute("""
+        UPDATE tracks SET is_instrumental = 1
+        WHERE has_lyrics = 0
+        AND (parakeet_transcription IS NULL OR LENGTH(parakeet_transcription) < 10)
+        AND is_instrumental = 0
+    """)
+    conn.execute("""
+        UPDATE tracks SET is_instrumental = 0
+        WHERE has_transcription = 1
+        AND LENGTH(parakeet_transcription) >= 10
+        AND is_instrumental = 1
+    """)
+    conn.commit()
+    instr = conn.execute("SELECT COUNT(*) FROM tracks WHERE is_instrumental=1").fetchone()[0]
+    log.info(f"  Instrumental tracks: {instr:,}")
+
+    conn.close()
+
+    elapsed = time.time() - t0
+    log.info(f"\nDone in {timedelta(seconds=int(elapsed))}")
+    log.info("Restart the server to load the updated indices.")
+
+
+if __name__ == "__main__":
+    update_all()

whisper_embeddings/riffusion_000013.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74b952d71827043bd9ab6c23e19068d0f0f4a6027f7c7ed4b7b766258584a013
+size 3489879

whisper_embeddings/sonauto_000001.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96bb2488c814d067a8144764cf79a25362eed8810bf366df2438ac4307e3064c
+size 4941349

whisper_embeddings/udio_000032.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a650987c9246b90f979e15afd12d4838bafee7f18ee99983ff2c99f2e00e6bc
+size 382643