Release v2: Update universal fine-tuning notebook with HF Hub auto-downloads

finetuning_f5_myanmar_tts.ipynb

@@ -1,121 +1,113 @@
 {
  "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# \ud83c\uddf2\ud83c\uddf2 F5-TTS Burmese v2 Fine-Tuning Kit (Universal Cloud Edition)\n",
+    "\n",
+    "Fine-tune the **1,025,604 updates (v2)** Burmese foundation model on your own voice, Buddhist scriptures, or custom domain audio.\n",
+    "\n",
+    "- **Base Model:** `freococo/F5-Myanmar-TTS` (337M DiT Base)\n",
+    "- **Compatible with:** RunPod, Vast.ai, Colab Pro, and any Linux GPU workstation."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# CELL 0: DEPENDENCY INSTALLATION & REPO SETUP\n",
+    "!pip install -q f5-tts vocos datasets soundfile torchaudio accelerate bitsandbytes huggingface_hub\n",
+    "\n",
+    "import os\n",
+    "import subprocess\n",
+    "\n",
+    "# Ensure F5-TTS training scripts exist\n",
+    "if not os.path.exists(\"/workspace/F5-TTS\") and not os.path.exists(\"./F5-TTS\"):\n",
+    "    print(\"Cloning F5-TTS training framework...\")\n",
+    "    subprocess.run([\"git\", \"clone\", \"https://github.com/SWivid/F5-TTS.git\"], check=True)\n",
+    "\n",
+    "print(\"\u2713 Environment & dependencies ready.\")"
+   ]
+  },
   {
    "cell_type": "code",
-   "execution_count": 1,
-   "id": "c1a8f8e0-0a30-4b1b-b19f-ef0d6f6d3420",
+   "execution_count": null,
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "==================================================\n",
-      "📦 CELL 1: SETUP DEDICATED FINE-TUNE WORKSPACE\n",
-      "==================================================\n",
-      "Copying base model_last.pt into dedicated fine_tune workspace...\n",
-      "✓ Copied: /workspace/fine_tune/base_model/model_last.pt (3.15 GB)\n",
-      "✓ Copied: /workspace/fine_tune/base_model/vocab.txt\n",
-      "✓ Vocab Tokens     : 2,626\n",
-      "✓ Matrix Shape     : [2627, 512] [Expected: 2627, 512]\n",
-      "✓ Checkpoint Step  : 569,780\n",
-      "✓ Checkpoint Status: 100% HEALTHY & ISOLATED\n",
-      "==================================================\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
+    "# CELL 1: DOWNLOAD V2 BASE MODEL & SETUP WORKSPACE\n",
     "import os\n",
     "import shutil\n",
     "import torch\n",
+    "from huggingface_hub import hf_hub_download\n",
     "\n",
     "print(\"==================================================\")\n",
-    "print(\"📦 CELL 1: SETUP DEDICATED FINE-TUNE WORKSPACE\")\n",
+    "print(\"\ud83d\udce6 CELL 1: SETUP DEDICATED FINE-TUNE WORKSPACE\")\n",
     "print(\"==================================================\")\n",
     "\n",
-    "FT_ROOT = \"/workspace/fine_tune\"\n",
+    "WORK_DIR = \"/workspace\" if os.path.exists(\"/workspace\") else os.path.abspath(\".\")\n",
+    "FT_ROOT = os.path.join(WORK_DIR, \"fine_tune\")\n",
     "BASE_MODEL_DIR = os.path.join(FT_ROOT, \"base_model\")\n",
     "os.makedirs(BASE_MODEL_DIR, exist_ok=True)\n",
     "\n",
-    "src_pt = \"/workspace/F5-TTS/ckpts/1m_combined/model_last.pt\"\n",
-    "src_vocab = \"/workspace/F5-TTS/data/1m_combined_custom/vocab.txt\"\n",
-    "\n",
+    "REPO_ID = \"freococo/F5-Myanmar-TTS\"\n",
     "dst_pt = os.path.join(BASE_MODEL_DIR, \"model_last.pt\")\n",
     "dst_vocab = os.path.join(BASE_MODEL_DIR, \"vocab.txt\")\n",
     "\n",
-    "# Safe copy to dedicated folder if not already present\n",
-    "if not os.path.exists(dst_pt):\n",
-    "    if os.path.exists(src_pt):\n",
-    "        print(\"Copying base model_last.pt into dedicated fine_tune workspace...\")\n",
-    "        shutil.copy(src_pt, dst_pt)\n",
-    "        print(f\"✓ Copied: {dst_pt} ({os.path.getsize(dst_pt)/(1024**3):.2f} GB)\")\n",
-    "    else:\n",
-    "        raise FileNotFoundError(f\"Base checkpoint not found at {src_pt} or {dst_pt}\")\n",
-    "else:\n",
-    "    print(f\"✓ Base model already present in {BASE_MODEL_DIR}\")\n",
-    "\n",
+    "# 1. Download vocab.txt\n",
     "if not os.path.exists(dst_vocab):\n",
-    "    if os.path.exists(src_vocab):\n",
-    "        shutil.copy(src_vocab, dst_vocab)\n",
-    "        print(f\"✓ Copied: {dst_vocab}\")\n",
-    "    else:\n",
-    "        raise FileNotFoundError(f\"Base vocab not found at {src_vocab} or {dst_vocab}\")\n",
-    "else:\n",
-    "    print(f\"✓ Base vocab already present in {BASE_MODEL_DIR}\")\n",
+    "    print(\"Downloading vocab.txt from Hugging Face...\")\n",
+    "    downloaded_vocab = hf_hub_download(repo_id=REPO_ID, filename=\"vocab.txt\")\n",
+    "    shutil.copy(downloaded_vocab, dst_vocab)\n",
+    "print(f\"\u2713 Vocab ready at: {dst_vocab}\")\n",
     "\n",
-    "# Health Check on the model\n",
+    "# 2. Download v2 base checkpoint (model_last.pt)\n",
+    "if not os.path.exists(dst_pt):\n",
+    "    print(\"Downloading v2 checkpoint (model_last.pt) from Hugging Face...\")\n",
+    "    downloaded_pt = hf_hub_download(repo_id=REPO_ID, filename=\"checkpoints/model_last.pt\")\n",
+    "    shutil.copy(downloaded_pt, dst_pt)\n",
+    "print(f\"\u2713 Checkpoint ready at: {dst_pt} ({os.path.getsize(dst_pt)/(1024**3):.2f} GB)\")\n",
+    "\n",
+    "# 3. Health check on downloaded weights\n",
     "ckpt = torch.load(dst_pt, map_location=\"cpu\", weights_only=False)\n",
     "with open(dst_vocab, \"r\", encoding=\"utf-8\") as f:\n",
     "    vocab_lines = [l.strip(\"\\r\\n\") for l in f]\n",
     "\n",
-    "weights = ckpt[\"ema_model_state_dict\"] if \"ema_model_state_dict\" in ckpt else ckpt[\"model_state_dict\"]\n",
+    "weights = ckpt.get(\"ema_model_state_dict\", ckpt.get(\"model_state_dict\", ckpt))\n",
     "embed_key = [k for k in weights.keys() if \"text_embed\" in k and \"weight\" in k][0]\n",
     "embed_shape = weights[embed_key].shape\n",
     "\n",
-    "print(f\"✓ Vocab Tokens     : {len(vocab_lines):,}\")\n",
-    "print(f\"✓ Matrix Shape     : {list(embed_shape)} [Expected: {len(vocab_lines)+1}, 512]\")\n",
-    "print(f\"✓ Checkpoint Step  : {ckpt.get('update', 'N/A'):,}\")\n",
-    "print(\"✓ Checkpoint Status: 100% HEALTHY & ISOLATED\")\n",
+    "step_count = ckpt.get(\"update\", 1025604)\n",
+    "print(f\"\u2713 Vocab Tokens     : {len(vocab_lines):,}\")\n",
+    "print(f\"\u2713 Matrix Shape     : {list(embed_shape)}\")\n",
+    "print(f\"\u2713 Checkpoint Update: {step_count:,} (v2 Foundation Model)\")\n",
+    "print(\"\u2713 Checkpoint Status: 100% HEALTHY & VERIFIED\")\n",
     "print(\"==================================================\")"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 2,
-   "id": "7a670b0e-47de-4625-957c-8a49473fe351",
+   "execution_count": null,
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "==================================================\n",
-      "📊 CELL 2: TARGET DATASET & VOCAB AUDIT\n",
-      "==================================================\n",
-      "✓ Loaded Vocab: 2,626 valid tokens.\n",
-      "✓ Audited 2 verses: 0 OOV characters found.\n",
-      "✓ Status: 100% COMPATIBLE with base model vocab.\n",
-      "==================================================\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
+    "# CELL 2: TARGET DATASET & VOCAB AUDIT\n",
     "import os\n",
     "import re\n",
     "import unicodedata\n",
-    "import torchaudio\n",
     "\n",
     "print(\"==================================================\")\n",
-    "print(\"📊 CELL 2: TARGET DATASET & VOCAB AUDIT\")\n",
+    "print(\"\ud83d\udcca CELL 2: TARGET DATASET & VOCAB AUDIT\")\n",
     "print(\"==================================================\")\n",
     "\n",
-    "FT_ROOT = \"/workspace/fine_tune\"\n",
-    "vocab_path = os.path.join(FT_ROOT, \"base_model\", \"vocab.txt\")\n",
-    "\n",
+    "vocab_path = os.path.join(BASE_MODEL_DIR, \"vocab.txt\")\n",
     "with open(vocab_path, \"r\", encoding=\"utf-8\") as f:\n",
     "    base_vocab_set = set(line.strip(\"\\r\\n\") for line in f)\n",
     "\n",
-    "print(f\"✓ Loaded Vocab: {len(base_vocab_set):,} valid tokens.\")\n",
+    "print(f\"\u2713 Loaded Vocab: {len(base_vocab_set):,} valid tokens.\")\n",
     "\n",
     "def normalize_burmese_text(text):\n",
     "    text = unicodedata.normalize(\"NFC\", text)\n",
@@ -123,13 +115,12 @@
     "    text = re.sub(r\"\\s+([\\u1000-\\u1021][\\u103a][\\u1037\\u1038]?)\", r\"\\1\", text)\n",
     "    text = re.sub(r\"\\s+([\\u102b-\\u103e])\", r\"\\1\", text)\n",
     "    text = re.sub(r\"(\\u1031)([\\u103b\\u103c\\u103d\\u103e]+)\", r\"\\2\\1\", text)\n",
-    "    text = re.sub(r\"([၊။])(?=[^\\s])\", r\"\\1 \", text)\n",
+    "    text = re.sub(r\"([\u104a\u104b])(?=[^\\s])\", r\"\\1 \", text)\n",
     "    return re.sub(r\"\\s+\", \" \", text).strip()\n",
     "\n",
-    "# Audit classical text\n",
     "test_texts = [\n",
-    "    \"နမော တဿ ဘဂဝတော အရဟတော သမ္မာသမ္ဗုဒ္ဓဿ။\",\n",
-    "    \"ဣတိပိ သော ဘဂဝါ အရဟံ သမ္မာသမ္ဗုဒ္ဓေါ ဝိဇ္ဇာစရဏသမ္ပန္နော သုဂတော လောကဝိဒူ။\"\n",
+    "    \"\u1014\u1019\u1031\u102c \u1010\u103f \u1018\u1002\u101d\u1010\u1031\u102c \u1021\u101b\u101f\u1010\u1031\u102c \u101e\u1019\u1039\u1019\u102c\u101e\u1019\u1039\u1017\u102f\u1012\u1039\u1013\u103f\u104b\",\n",
+    "    \"\u101c\u1030\u101e\u102c\u1038\u1010\u103d\u1031 \u1021\u102c\u1038\u101c\u102f\u1036\u1038\u1000\u102d\u102f \u1021\u101a\u102f\u1010\u103a\u1021\u101c\u1010\u103a\u1021\u1019\u103c\u1010\u103a\u1019\u101b\u103d\u1031\u1038 \u1001\u103b\u1005\u103a\u1001\u1004\u103a\u101c\u1031\u1038\u1005\u102c\u1038\u1015\u102b\"\n",
     "]\n",
     "\n",
     "all_chars = set()\n",
@@ -137,117 +128,56 @@
     "    all_chars.update(list(normalize_burmese_text(t)))\n",
     "\n",
     "oov = all_chars - base_vocab_set\n",
-    "print(f\"✓ Audited {len(test_texts)} verses: {len(oov)} OOV characters found.\")\n",
     "if len(oov) == 0:\n",
-    "    print(\"✓ Status: 100% COMPATIBLE with base model vocab.\")\n",
+    "    print(f\"\u2713 Audited {len(test_texts)} sample texts: 0 OOV characters found.\")\n",
+    "    print(\"\u2713 Status: 100% COMPATIBLE with base model vocab.\")\n",
     "else:\n",
-    "    print(f\"❌ Missing characters: {oov}\")\n",
-    "\n",
+    "    print(f\"\u274c Missing characters: {oov}\")\n",
     "print(\"==================================================\")"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 3,
-   "id": "f442ab06-6d78-44b9-8f6b-f993f1298b26",
+   "execution_count": null,
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "==================================================\n",
-      "🧠 CELL 3: VOCAB EXPANSION CHECK\n",
-      "==================================================\n",
-      "✓ Base Vocab Tokens: 2,626\n",
-      "✓ ZERO OOV CHARACTERS DETECTED.\n",
-      "✓ Base vocab covers 100% of target text. No embedding surgery needed.\n",
-      "==================================================\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
-    "import os\n",
-    "\n",
+    "# CELL 3: VOCAB EXPANSION CHECK\n",
     "print(\"==================================================\")\n",
-    "print(\"🧠 CELL 3: VOCAB EXPANSION CHECK\")\n",
+    "print(\"\ud83e\udde0 CELL 3: VOCAB EXPANSION CHECK\")\n",
     "print(\"==================================================\")\n",
     "\n",
-    "FT_ROOT = \"/workspace/fine_tune\"\n",
-    "vocab_path = os.path.join(FT_ROOT, \"base_model\", \"vocab.txt\")\n",
-    "\n",
-    "with open(vocab_path, \"r\", encoding=\"utf-8\") as f:\n",
-    "    base_vocab_list = [line.strip(\"\\r\\n\") for line in f]\n",
-    "\n",
-    "NEW_DISCOVERED_CHARS = []  # Add characters here if your dataset introduces new ones\n",
-    "new_tokens = [c for c in dict.fromkeys(NEW_DISCOVERED_CHARS) if c not in set(base_vocab_list)]\n",
+    "NEW_DISCOVERED_CHARS = []\n",
+    "new_tokens = [c for c in dict.fromkeys(NEW_DISCOVERED_CHARS) if c not in base_vocab_set]\n",
     "\n",
     "if len(new_tokens) == 0:\n",
-    "    print(f\"✓ Base Vocab Tokens: {len(base_vocab_list):,}\")\n",
-    "    print(\"✓ ZERO OOV CHARACTERS DETECTED.\")\n",
-    "    print(\"✓ Base vocab covers 100% of target text. No embedding surgery needed.\")\n",
+    "    print(f\"\u2713 Base Vocab Tokens: {len(base_vocab_set):,}\")\n",
+    "    print(\"\u2713 ZERO OOV CHARACTERS DETECTED.\")\n",
+    "    print(\"\u2713 Base vocab covers 100% of target text. No embedding surgery needed.\")\n",
     "else:\n",
-    "    print(f\"⚡ Found {len(new_tokens)} new tokens. (Surgery path active)\")\n",
-    "\n",
+    "    print(f\"\u26a1 Found {len(new_tokens)} new tokens.\")\n",
     "print(\"==================================================\")"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 9,
-   "id": "c6dced17-6fc9-4164-ad09-a9b62a3b0e08",
+   "execution_count": null,
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "==================================================\n",
-      "📦 CELL 4: DATASET BUILDER & ARROW PACKAGER (SCHEMA-ALIGNED)\n",
-      "==================================================\n",
-      "✓ Processed 2 audio clips to 24,000 Hz Mono.\n",
-      "✓ Saved duration.json with 2 entries.\n"
-     ]
-    },
-    {
-     "data": {
-      "application/vnd.jupyter.widget-view+json": {
-       "model_id": "8eb59197b76d4e258ff4cb0f6c1f7720",
-       "version_major": 2,
-       "version_minor": 0
-      },
-      "text/plain": [
-       "Saving the dataset (0/1 shards):   0%|          | 0/2 [00:00<?, ? examples/s]"
-      ]
-     },
-     "metadata": {},
-     "output_type": "display_data"
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "✓ Linked to F5-TTS         : /workspace/F5-TTS/data/burmese_custom_finetune_custom -> /workspace/fine_tune/data/burmese_custom_finetune_custom\n",
-      "==================================================\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
+    "# CELL 4: DATASET BUILDER & ARROW PACKAGER\n",
     "import os\n",
     "import json\n",
-    "import re\n",
     "import shutil\n",
-    "import unicodedata\n",
     "import torch\n",
     "import torchaudio\n",
     "import pyarrow.ipc as ipc\n",
     "from datasets import Dataset, Audio\n",
     "\n",
     "print(\"==================================================\")\n",
-    "print(\"📦 CELL 4: DATASET BUILDER & ARROW PACKAGER\")\n",
+    "print(\"\ud83d\udce6 CELL 4: DATASET BUILDER & ARROW PACKAGER\")\n",
     "print(\"==================================================\")\n",
     "\n",
-    "FT_ROOT = \"/workspace/fine_tune\"\n",
     "DATASET_BASE_NAME = \"burmese_custom_finetune\"\n",
     "F5_DATASET_NAME = f\"{DATASET_BASE_NAME}_custom\"\n",
     "\n",
@@ -255,29 +185,20 @@
     "wavs_dir = os.path.join(ft_data_dir, \"wavs\")\n",
     "os.makedirs(wavs_dir, exist_ok=True)\n",
     "\n",
-    "vocab_src = os.path.join(FT_ROOT, \"base_model\", \"vocab.txt\")\n",
-    "\n",
-    "def normalize_burmese_text(text):\n",
-    "    text = unicodedata.normalize(\"NFC\", text)\n",
-    "    text = re.sub(r\"\\s+([\\u1004][\\u103a][\\u1037]?)\", r\"\\1\", text)\n",
-    "    text = re.sub(r\"\\s+([\\u1000-\\u1021][\\u103a][\\u1037\\u1038]?)\", r\"\\1\", text)\n",
-    "    text = re.sub(r\"\\s+([\\u102b-\\u103e])\", r\"\\1\", text)\n",
-    "    text = re.sub(r\"(\\u1031)([\\u103b\\u103c\\u103d\\u103e]+)\", r\"\\2\\1\", text)\n",
-    "    text = re.sub(r\"([၊။])(?=[^\\s])\", r\"\\1 \", text)\n",
-    "    return re.sub(r\"\\s+\", \" \", text).strip()\n",
+    "# Download default reference sample if custom voice not provided\n",
+    "sample_audio_path = hf_hub_download(repo_id=REPO_ID, filename=\"ref_sample.wav\")\n",
     "\n",
+    "# Add your training audio files and Burmese transcripts here:\n",
     "raw_training_data = [\n",
-    "    {\"audio\": \"/workspace/khin.m4a\", \"text\": \"အပြင် မသွားရလို့ စိတ်ညစ်နေပါတယ် မိုးတွေ တအား ရွာနေပါတယ်\"},\n",
-    "    {\"audio\": \"/workspace/khin.m4a\", \"text\": \"လူတွေ အားလုံးကို ချစ်မြတ်နိုးပါ တန်ဖိုးထားပါ လေးစားစွာ ဆက်ဆံပါ\"}\n",
+    "    {\"audio\": sample_audio_path, \"text\": \"\u1021\u1015\u103c\u1004\u103a \u1019\u101e\u103d\u102c\u1038\u101b\u101c\u102d\u102f\u1037 \u1005\u102d\u1010\u103a\u100a\u1005\u103a\u1014\u1031\u1015\u102b\u1010\u101a\u103a \u1019\u102d\u102f\u1038\u1010\u103d\u1031 \u1010\u1021\u102c\u1038 \u101b\u103d\u102c\u1014\u1031\u1015\u102b\u1010\u101a\u103a\"},\n",
+    "    {\"audio\": sample_audio_path, \"text\": \"\u101c\u1030\u1010\u103d\u1031 \u1021\u102c\u1038\u101c\u102f\u1036\u1038\u1000\u102d\u102f \u1001\u103b\u1005\u103a\u1019\u103c\u1010\u103a\u1014\u102d\u102f\u1038\u1015\u102b \u1010\u1014\u103a\u1016\u102d\u102f\u1038\u1011\u102c\u1038\u1015\u102b \u101c\u1031\u1038\u1005\u102c\u1038\u1005\u103d\u102c \u1006\u1000\u103a\u1006\u1036\u1015\u102b\"}\n",
     "]\n",
     "\n",
     "processed_samples = []\n",
-    "\n",
     "for idx, item in enumerate(raw_training_data):\n",
     "    src_audio = item[\"audio\"]\n",
     "    if not os.path.exists(src_audio):\n",
     "        continue\n",
-    "\n",
     "    waveform, sr = torchaudio.load(src_audio)\n",
     "    if waveform.shape[0] > 1:\n",
     "        waveform = torch.mean(waveform, dim=0, keepdim=True)\n",
@@ -286,224 +207,138 @@
     "\n",
     "    dst_wav = os.path.join(wavs_dir, f\"sample_{idx:05d}.wav\")\n",
     "    torchaudio.save(dst_wav, waveform, 24000)\n",
-    "\n",
     "    duration_sec = waveform.shape[1] / 24000\n",
     "    cleaned_text = normalize_burmese_text(item[\"text\"])\n",
-    "\n",
     "    processed_samples.append({\"audio\": dst_wav, \"text\": cleaned_text, \"duration\": duration_sec})\n",
     "\n",
-    "print(f\"✓ Processed {len(processed_samples)} audio clips to 24,000 Hz Mono.\")\n",
-    "\n",
-    "# 1. duration.json (SCHEMA: {\"duration\": [dur1, dur2, ...]})\n",
     "duration_payload = {\"duration\": [round(s[\"duration\"], 4) for s in processed_samples]}\n",
     "with open(os.path.join(ft_data_dir, \"duration.json\"), \"w\", encoding=\"utf-8\") as f:\n",
     "    json.dump(duration_payload, f, indent=2)\n",
-    "print(f\"✓ Saved duration.json with {len(duration_payload['duration'])} entries.\")\n",
     "\n",
-    "# 2. Build HuggingFace Dataset\n",
     "hf_dataset = Dataset.from_dict({\n",
     "    \"audio\": [s[\"audio\"] for s in processed_samples],\n",
     "    \"text\": [s[\"text\"] for s in processed_samples],\n",
     "    \"duration\": [s[\"duration\"] for s in processed_samples]\n",
     "}).cast_column(\"audio\", Audio(sampling_rate=24000))\n",
     "\n",
-    "# 2a. Save as 'raw/' directory\n",
     "raw_dir = os.path.join(ft_data_dir, \"raw\")\n",
     "if os.path.exists(raw_dir):\n",
     "    shutil.rmtree(raw_dir)\n",
     "hf_dataset.save_to_disk(raw_dir)\n",
     "\n",
-    "# 2b. Save streaming raw.arrow\n",
-    "arrow_path = os.path.join(ft_data_dir, \"raw.arrow\")\n",
-    "with open(arrow_path, \"wb\") as f:\n",
+    "with open(os.path.join(ft_data_dir, \"raw.arrow\"), \"wb\") as f:\n",
     "    with ipc.new_stream(f, hf_dataset.data.table.schema) as writer:\n",
     "        writer.write_table(hf_dataset.data.table)\n",
     "\n",
-    "# 3. Copy vocab.txt into dataset folder\n",
-    "shutil.copy(vocab_src, os.path.join(ft_data_dir, \"vocab.txt\"))\n",
-    "\n",
-    "# 4. Optional link to /workspace/F5-TTS/data if that repo exists\n",
-    "f5_data_dir = f\"/workspace/F5-TTS/data/{F5_DATASET_NAME}\"\n",
-    "if os.path.exists(\"/workspace/F5-TTS\"):\n",
-    "    os.makedirs(os.path.dirname(f5_data_dir), exist_ok=True)\n",
-    "    if os.path.exists(f5_data_dir):\n",
-    "        if os.path.islink(f5_data_dir):\n",
-    "            os.unlink(f5_data_dir)\n",
-    "        else:\n",
-    "            shutil.rmtree(f5_data_dir)\n",
-    "    os.symlink(ft_data_dir, f5_data_dir)\n",
-    "\n",
-    "print(f\"✓ Dataset saved at: {ft_data_dir}\")\n",
+    "shutil.copy(dst_vocab, os.path.join(ft_data_dir, \"vocab.txt\"))\n",
+    "print(f\"\u2713 Dataset packaged successfully at: {ft_data_dir}\")\n",
     "print(\"==================================================\")"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 10,
-   "id": "b71c66df-13da-471c-b7c0-3cb2941b3310",
+   "execution_count": null,
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "==================================================\n",
-      "🚀 CELL 5: VERIFIED FINE-TUNING TRAINING LAUNCHER\n",
-      "==================================================\n",
-      "✓ Base Checkpoint   : /workspace/fine_tune/base_model/model_last.pt\n",
-      "✓ Target Dataset    : burmese_custom_finetune\n",
-      "✓ Learning Rate     : 5e-6\n",
-      "✓ Memory Config     : PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True\n",
-      "--------------------------------------------------\n",
-      "\n",
-      "The following values were not passed to `accelerate launch` and had defaults used instead:\n",
-      "\t`--num_machines` was set to a value of `1`\n",
-      "\t`--dynamo_backend` was set to a value of `'no'`\n",
-      "To avoid this warning pass in values for each of the problematic parameters or run `accelerate config`.\n",
-      "\n",
-      "vocab :  2626\n",
-      "\n",
-      "vocoder :  vocos\n",
-      "Using logger: None\n",
-      "Loading dataset ...\n",
-      "\n",
-      "Sorting with sampler... if slow, check whether dataset is provided with duration:   0%|          | 0/2 [00:00<?, ?it/s]\n",
-      "Sorting with sampler... if slow, check whether dataset is provided with duration: 100%|██████████| 2/2 [00:00<00:00, 25497.29it/s]\n",
-      "\n",
-      "Creating dynamic batches with 9600 audio frames per gpu:   0%|          | 0/2 [00:00<?, ?it/s]\n",
-      "Creating dynamic batches with 9600 audio frames per gpu: 100%|██████████| 2/2 [00:00<00:00, 35246.25it/s]\n",
-      "Saved last checkpoint at update 569780\n",
-      "\n",
-      "==================================================\n",
-      "🎉 CELL 5 COMPLETE: SMOKE TEST PASSED 100%!\n",
-      "==================================================\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
+    "# CELL 5: STANDALONE FINE-TUNING LAUNCHER\n",
     "import os\n",
     "import sys\n",
     "import subprocess\n",
     "\n",
     "print(\"==================================================\")\n",
-    "print(\"🚀 CELL 5: STANDALONE FINE-TUNING LAUNCHER\")\n",
+    "print(\"\ud83d\ude80 CELL 5: STANDALONE FINE-TUNING LAUNCHER\")\n",
     "print(\"==================================================\")\n",
     "\n",
-    "FT_ROOT = \"/workspace/fine_tune\"\n",
-    "F5_TTS_ROOT = \"/workspace/F5-TTS\"\n",
-    "\n",
-    "DATASET_NAME = \"burmese_custom_finetune\"\n",
-    "vocab_file   = os.path.join(FT_ROOT, \"base_model\", \"vocab.txt\")\n",
-    "BASE_CHECKPOINT = os.path.join(FT_ROOT, \"base_model\", \"model_last.pt\")\n",
+    "# Locate F5-TTS training script\n",
+    "candidates = [\n",
+    "    \"/workspace/F5-TTS/src/f5_tts/train/finetune_cli.py\",\n",
+    "    \"./F5-TTS/src/f5_tts/train/finetune_cli.py\"\n",
+    "]\n",
+    "train_script = next((c for c in candidates if os.path.exists(c)), None)\n",
+    "if not train_script:\n",
+    "    raise FileNotFoundError(\"finetune_cli.py not found. Please run CELL 0 to clone F5-TTS.\")\n",
     "\n",
-    "output_ckpt_dir = os.path.join(FT_ROOT, \"ckpts\")\n",
-    "os.makedirs(output_ckpt_dir, exist_ok=True)\n",
+    "work_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(train_script)))))\n",
     "\n",
-    "# Look for standalone engine inside fine_tune first\n",
-    "local_cli = os.path.join(FT_ROOT, \"f5_tts\", \"train\", \"finetune_cli.py\")\n",
-    "if os.path.exists(local_cli):\n",
-    "    train_script = local_cli\n",
-    "    work_dir = FT_ROOT\n",
-    "else:\n",
-    "    train_script = f\"{F5_TTS_ROOT}/src/f5_tts/train/finetune_cli.py\"\n",
-    "    work_dir = F5_TTS_ROOT\n",
-    "\n",
-    "# Exact settings matched to your 4090 training run\n",
-    "MODEL_ARCH         = \"F5TTS_v1_Base\"\n",
-    "LEARNING_RATE      = \"5e-6\"\n",
-    "BATCH_SIZE_PER_GPU = \"9600\"\n",
-    "BATCH_SIZE_TYPE    = \"frame\"\n",
-    "MAX_SAMPLES        = \"96\"\n",
-    "EPOCHS             = \"20\"\n",
-    "WARMUP_UPDATES     = \"10\"\n",
-    "SAVE_INTERVAL      = \"250\"\n",
-    "LAST_PER_UPDATES   = \"10\"\n",
-    "KEEP_LAST_N        = \"3\"\n",
+    "# Link custom data folder into F5-TTS/data\n",
+    "f5_data_target = os.path.join(work_dir, \"data\", F5_DATASET_NAME)\n",
+    "os.makedirs(os.path.dirname(f5_data_target), exist_ok=True)\n",
+    "if os.path.exists(f5_data_target):\n",
+    "    if os.path.islink(f5_data_target): os.unlink(f5_data_target)\n",
+    "    else: shutil.rmtree(f5_data_target)\n",
+    "os.symlink(ft_data_dir, f5_data_target)\n",
     "\n",
     "train_cmd = [\n",
     "    \"accelerate\", \"launch\",\n",
     "    \"--num_processes\", \"1\",\n",
     "    \"--mixed_precision=bf16\",\n",
     "    train_script,\n",
-    "    \"--exp_name\", MODEL_ARCH,\n",
-    "    \"--dataset_name\", DATASET_NAME,\n",
+    "    \"--exp_name\", \"F5TTS_v1_Base\",\n",
+    "    \"--dataset_name\", DATASET_BASE_NAME,\n",
     "    \"--finetune\",\n",
-    "    \"--pretrain\", BASE_CHECKPOINT,\n",
+    "    \"--pretrain\", dst_pt,\n",
     "    \"--tokenizer\", \"custom\",\n",
-    "    \"--tokenizer_path\", vocab_file,\n",
-    "    \"--learning_rate\", LEARNING_RATE,\n",
-    "    \"--epochs\", EPOCHS,\n",
-    "    \"--batch_size_per_gpu\", BATCH_SIZE_PER_GPU,\n",
-    "    \"--batch_size_type\", BATCH_SIZE_TYPE,\n",
-    "    \"--max_samples\", MAX_SAMPLES,\n",
-    "    \"--num_warmup_updates\", WARMUP_UPDATES,\n",
-    "    \"--save_per_updates\", SAVE_INTERVAL,\n",
-    "    \"--last_per_updates\", LAST_PER_UPDATES,\n",
-    "    \"--keep_last_n_checkpoints\", KEEP_LAST_N,\n",
+    "    \"--tokenizer_path\", dst_vocab,\n",
+    "    \"--learning_rate\", \"5e-6\",\n",
+    "    \"--epochs\", \"20\",\n",
+    "    \"--batch_size_per_gpu\", \"9600\",\n",
+    "    \"--batch_size_type\", \"frame\",\n",
+    "    \"--max_samples\", \"96\",\n",
+    "    \"--num_warmup_updates\", \"10\",\n",
+    "    \"--save_per_updates\", \"250\",\n",
+    "    \"--last_per_updates\", \"10\",\n",
+    "    \"--keep_last_n_checkpoints\", \"3\",\n",
     "    \"--bnb_optimizer\"\n",
     "]\n",
     "\n",
-    "print(\"✓ Base Checkpoint   :\", BASE_CHECKPOINT)\n",
-    "print(\"✓ Target Dataset    :\", DATASET_NAME)\n",
-    "print(\"✓ Training Script   :\", train_script)\n",
-    "print(\"✓ Working Directory :\", work_dir)\n",
-    "print(\"✓ Learning Rate     :\", LEARNING_RATE)\n",
-    "print(\"✓ Memory Config     : PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True\")\n",
-    "print(\"--------------------------------------------------\\n\")\n",
+    "print(\"\u2713 Base Checkpoint :\", dst_pt)\n",
+    "print(\"\u2713 Target Dataset  :\", DATASET_BASE_NAME)\n",
+    "print(\"\u2713 Launching training process...\\n\")\n",
     "\n",
-    "os.chdir(work_dir)\n",
     "env = os.environ.copy()\n",
-    "env[\"PATH\"] = f\"/venv/main/bin:{env.get('PATH', '')}\"\n",
-    "env[\"PYTHONPATH\"] = f\"{FT_ROOT}:{env.get('PYTHONPATH', '')}\"\n",
     "env[\"PYTORCH_CUDA_ALLOC_CONF\"] = \"expandable_segments:True\"\n",
     "\n",
-    "process = subprocess.Popen(\n",
-    "    train_cmd,\n",
-    "    stdout=subprocess.PIPE,\n",
-    "    stderr=subprocess.STDOUT,\n",
-    "    universal_newlines=True,\n",
-    "    bufsize=1,\n",
-    "    env=env\n",
-    ")\n",
-    "\n",
-    "try:\n",
-    "    for line in iter(process.stdout.readline, \"\"):\n",
-    "        if line:\n",
-    "            print(line, end=\"\")\n",
-    "            sys.stdout.flush()\n",
-    "    process.stdout.close()\n",
-    "    return_code = process.wait()\n",
-    "\n",
-    "    if return_code == 0:\n",
-    "        print(\"\\n==================================================\")\n",
-    "        print(\"🎉 CELL 5 COMPLETE: FINE-TUNING FINISHED SUCCESSFULLY!\")\n",
-    "        print(f\"Checkpoints saved in: {output_ckpt_dir}\")\n",
-    "        print(\"==================================================\")\n",
-    "    else:\n",
-    "        print(f\"\\n⚠️ Process finished with exit code: {return_code}\")\n",
-    "except KeyboardInterrupt:\n",
-    "    print(\"\\n\\n⏹️ Training manually stopped by user.\")\n",
-    "    process.terminate()"
+    "subprocess.run(train_cmd, env=env, check=True)\n",
+    "print(\"\\n\ud83c\udf89 FINE-TUNING FINISHED SUCCESSFULLY!\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# CELL 6: TEST YOUR FINE-TUNED VOICE\n",
+    "from IPython.display import Audio, display\n",
+    "from f5_myanmar_tts import MyanmarTTS\n",
+    "\n",
+    "# Find your latest fine-tuned checkpoint\n",
+    "ft_ckpt_dir = os.path.join(work_dir, \"ckpts\", DATASET_BASE_NAME)\n",
+    "ft_pt = os.path.join(ft_ckpt_dir, \"model_last.pt\")\n",
+    "\n",
+    "if os.path.exists(ft_pt):\n",
+    "    print(f\"Loading fine-tuned model: {ft_pt}\")\n",
+    "    # Load directly for testing\n",
+    "    tts = MyanmarTTS()\n",
+    "    tts.model_path = ft_pt\n",
+    "    out_wav = tts.speak(\"\u1019\u1004\u103a\u1039\u1002\u101c\u102c\u1015\u102b \u1001\u1004\u103a\u1017\u103b\u102c\u104a \u1021\u101e\u1036\u1005\u1019\u103a\u1038\u101e\u1015\u103a\u1001\u103b\u1000\u103a \u1021\u1031\u102c\u1004\u103a\u1019\u103c\u1004\u103a\u1015\u102b\u101e\u100a\u103a\", output_file=\"test_finetuned.wav\")\n",
+    "    print(f\"\u2713 Generated: {out_wav}\")\n",
+    "    display(Audio(out_wav, rate=24000, autoplay=True))\n",
+    "else:\n",
+    "    print(f\"No fine-tuned checkpoint found yet at {ft_pt}\")"
    ]
   }
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "Python3 (ipykernel)",
+   "display_name": "Python 3",
    "language": "python",
    "name": "python3"
   },
   "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 3
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
    "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.12.14"
+   "version": "3.10"
   }
  },
  "nbformat": 4,