Fix Cell 14: replace evaluate.load('wer') with jiwer directly

evaluate==0.4.2 uses use_auth_token which was removed from
huggingface-hub>=1.0 -- raises AttributeError at runtime.
Replaced with jiwer.wer() + jiwer.Compose transforms, which needs
no network call and is already installed.

Also removed evaluate from Cell 2 install list.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

notebooks/kaggle_master_trainer.ipynb

@@ -33,7 +33,7 @@
    "id": "cell-install",
    "metadata": {},
    "outputs": [],
-   "source": "# -- Cell 2: Install dependencies + fix torchvision CUDA mismatch -------------\n# Kaggle ships torchvision built against CUDA 12.8 but torch uses CUDA 13.0.\n# Any import touching torchvision (WhisperProcessor, WhisperForConditionalGeneration)\n# raises RuntimeError. The fix requires reinstalling torchvision AND restarting\n# the kernel so the new version is loaded from scratch.\n#\n# This cell uses a sentinel file so it only restarts once:\n#   First run  -> installs correct torchvision -> restarts kernel\n#   Second run -> sentinel exists -> skips restart -> continues normally\n\nimport subprocess, sys, os\nfrom pathlib import Path\n\nSENTINEL = Path('/kaggle/working/.torchvision_fixed')\n\nif not SENTINEL.exists():\n    import torch\n    cuda_tag = 'cu' + torch.version.cuda.replace('.', '')[:3]   # e.g. cu130\n    print(f'Fixing torchvision for CUDA {cuda_tag} (will restart kernel) ...')\n\n    subprocess.check_call([\n        sys.executable, '-m', 'pip', 'install', '-q', '--upgrade',\n        'torchvision',\n        '--index-url', f'https://download.pytorch.org/whl/{cuda_tag}',\n    ])\n\n    pkgs = [\n        'transformers==5.5.0', 'datasets==4.8.4', 'accelerate==1.13.0',\n        'evaluate==0.4.2', 'huggingface-hub==1.9.0', 'peft==0.18.1',\n        'bitsandbytes==0.49.2', 'librosa==0.10.2', 'soundfile==0.12.1',\n        'jiwer==3.0.4', 'numpy==2.2.4', 'scipy==1.15.2',\n    ]\n    subprocess.check_call([sys.executable, '-m', 'pip', 'install', '-q', *pkgs])\n\n    SENTINEL.touch()\n    print('Packages installed. Restarting kernel to load new torchvision ...')\n    import IPython\n    IPython.Application.instance().kernel.do_shutdown(restart=True)\n\nelse:\n    print('torchvision already fixed (sentinel exists) -- skipping restart.')\n    import torch\n    print(f'torch     : {torch.__version__}')\n    print(f'CUDA      : {torch.version.cuda}')\n    print(f'GPU       : {torch.cuda.get_device_name(0) if torch.cuda.is_available() else \"none\"}')\n    print('Ready.')"
+   "source": "# -- Cell 2: Install dependencies + fix torchvision CUDA mismatch -------------\nimport subprocess, sys\nfrom pathlib import Path\n\nSENTINEL = Path('/kaggle/working/.torchvision_fixed')\n\nif not SENTINEL.exists():\n    import torch\n    cuda_tag = 'cu' + torch.version.cuda.replace('.', '')[:3]\n    print(f'Fixing torchvision for CUDA {cuda_tag} (will restart kernel) ...')\n    subprocess.check_call([\n        sys.executable, '-m', 'pip', 'install', '-q', '--upgrade',\n        'torchvision', '--index-url', f'https://download.pytorch.org/whl/{cuda_tag}',\n    ])\n    pkgs = [\n        'transformers==5.5.0', 'datasets==4.8.4', 'accelerate==1.13.0',\n        'huggingface-hub==1.9.0', 'peft==0.18.1',\n        'bitsandbytes==0.49.2', 'librosa==0.10.2', 'soundfile==0.12.1',\n        'jiwer==3.0.4', 'numpy==2.2.4', 'scipy==1.15.2',\n        # NOTE: evaluate is intentionally excluded -- evaluate.load('wer') breaks\n        # with huggingface-hub>=1.0 (use_auth_token removed). We use jiwer directly.\n    ]\n    subprocess.check_call([sys.executable, '-m', 'pip', 'install', '-q', *pkgs])\n    SENTINEL.touch()\n    print('Restarting kernel to load fixed torchvision ...')\n    import IPython\n    IPython.Application.instance().kernel.do_shutdown(restart=True)\nelse:\n    import torch\n    print('torchvision already fixed -- skipping restart.')\n    print(f'torch : {torch.__version__}  CUDA : {torch.version.cuda}')\n    print(f'GPU   : {torch.cuda.get_device_name(0) if torch.cuda.is_available() else \"none\"}')"
   },
   {
    "cell_type": "code",
@@ -149,7 +149,7 @@
    "id": "cell-collator",
    "metadata": {},
    "outputs": [],
-   "source": "# ── Cell 14: Data collator + WER metric ──────────────────────────────────────\nimport evaluate\nfrom dataclasses import dataclass\nfrom typing import Any, Dict, List\n\nwer_metric = evaluate.load('wer')\n\n\n@dataclass\nclass DataCollatorSpeechSeq2SeqWithPadding:\n    \"\"\"\n    Pads input_features (mel spectrograms) and labels independently.\n    Labels use -100 for padding so CrossEntropyLoss ignores them.\n    Strips the BOS token from label sequences (Whisper convention).\n    \"\"\"\n    processor: Any\n\n    def __call__(self, features: List[Dict]) -> Dict:\n        # Pad mel features\n        input_feats = [{'input_features': f['input_features']} for f in features]\n        batch = self.processor.feature_extractor.pad(input_feats, return_tensors='pt')\n\n        # Pad labels\n        label_feats  = [{'input_ids': f['labels']} for f in features]\n        labels_batch = self.processor.tokenizer.pad(label_feats, return_tensors='pt')\n        labels       = labels_batch['input_ids'].masked_fill(\n            labels_batch.attention_mask.ne(1), -100\n        )\n        # Remove leading BOS token (Whisper adds it during generation, not training)\n        if (labels[:, 0] == self.processor.tokenizer.bos_token_id).all().item():\n            labels = labels[:, 1:]\n        batch['labels'] = labels\n        return batch\n\n\ndef compute_metrics(pred):\n    \"\"\"WER on the eval set (lower is better).\"\"\"\n    pred_ids  = pred.predictions\n    label_ids = pred.label_ids\n    # Replace -100 padding back to pad_token_id for decoding\n    label_ids[label_ids == -100] = processor.tokenizer.pad_token_id\n\n    pred_str  = processor.tokenizer.batch_decode(pred_ids,  skip_special_tokens=True)\n    label_str = processor.tokenizer.batch_decode(label_ids, skip_special_tokens=True)\n    wer = wer_metric.compute(predictions=pred_str, references=label_str)\n    return {'wer': round(wer, 4)}\n\n\ncollator = DataCollatorSpeechSeq2SeqWithPadding(processor=processor)\nprint('✅ Collator and WER metric ready')"
+   "source": "# -- Cell 14: Data collator + WER metric --------------------------------------\n# evaluate.load('wer') breaks with huggingface-hub>=1.0 (use_auth_token removed).\n# Use jiwer directly instead -- same result, no network call needed.\nimport jiwer\nfrom dataclasses import dataclass\nfrom typing import Any, Dict, List\n\ntransform = jiwer.Compose([\n    jiwer.ToLowerCase(),\n    jiwer.RemoveMultipleSpaces(),\n    jiwer.Strip(),\n    jiwer.RemovePunctuation(),\n    jiwer.ReduceToListOfListOfWords(),\n])\n\n\n@dataclass\nclass DataCollatorSpeechSeq2SeqWithPadding:\n    processor: Any\n\n    def __call__(self, features: List[Dict]) -> Dict:\n        import torch\n        # Pad mel features\n        input_feats  = [{'input_features': f['input_features']} for f in features]\n        batch        = self.processor.feature_extractor.pad(input_feats, return_tensors='pt')\n\n        # Pad labels, replace padding with -100\n        label_feats  = [{'input_ids': f['labels']} for f in features]\n        labels_batch = self.processor.tokenizer.pad(label_feats, return_tensors='pt')\n        labels       = labels_batch['input_ids'].masked_fill(\n            labels_batch.attention_mask.ne(1), -100\n        )\n        # Strip leading BOS token\n        if (labels[:, 0] == self.processor.tokenizer.bos_token_id).all().item():\n            labels = labels[:, 1:]\n        batch['labels'] = labels\n        return batch\n\n\ndef compute_metrics(pred):\n    pred_ids  = pred.predictions\n    label_ids = pred.label_ids\n    label_ids[label_ids == -100] = processor.tokenizer.pad_token_id\n\n    pred_str  = processor.tokenizer.batch_decode(pred_ids,  skip_special_tokens=True)\n    label_str = processor.tokenizer.batch_decode(label_ids, skip_special_tokens=True)\n\n    wer = jiwer.wer(label_str, pred_str,\n                    hypothesis_transform=transform,\n                    reference_transform=transform)\n    return {'wer': round(wer, 4)}\n\n\ncollator = DataCollatorSpeechSeq2SeqWithPadding(processor=processor)\nprint('Collator and WER metric ready')"
   },
   {
    "cell_type": "markdown",