anyscale · Aydin-ab · Jul 9, 2026 · Jul 9, 2026 · Jul 9, 2026 · Jul 9, 2026
@@ -303,7 +303,7 @@
   owner_team: llm
   dir: templates/audio-dataset-curation-llm-judge
   cluster_env:
-    image_uri: anyscale/ray-llm:2.55.1-py311-cu128
+    image_uri: anyscale/ray-llm:2.56.0-py312-cu130
   compute_config:
     AWS: configs/audio-dataset-curation-llm-judge/aws.yaml
     GCP: configs/audio-dataset-curation-llm-judge/gce.yaml

diff --git a/dependencies/template.depsets.yaml b/dependencies/template.depsets.yaml
@@ -149,12 +149,12 @@ depsets:
       - templates/audio-dataset-curation-llm-judge/requirements.txt
     output: templates/audio-dataset-curation-llm-judge/python_depset.lock
     append_flags:
-      - --index https://download.pytorch.org/whl/cu128
+      - --index https://download.pytorch.org/whl/cu130
       - --python-version=${PYTHON_VERSION}
       - --python-platform=x86_64-manylinux_2_31
       - --unsafe-package ray
     build_arg_sets:
-      - ray2551_py311_cu128
+      - ray2560_py312_cu130
 
   - name: multi_agent_a2a_depset_${RAY_VERSION}_${PYTHON_VERSION}_${CUDA_VARIANT}
     operation: expand

diff --git a/templates/audio-dataset-curation-llm-judge/README.ipynb b/templates/audio-dataset-curation-llm-judge/README.ipynb
@@ -7,11 +7,11 @@
     "# Audio batch inference\n",
     "\n",
     "<div align=\"left\">\n",
-    "  <a target=\"_blank\" href=\"https://console.anyscale.com/template-preview/audio-dataset-curation-llm-judge\"><img src=\"https://img.shields.io/badge/🚀 Run_on-Anyscale-9hf\"></a>&nbsp;\n",
+    "  <a target=\"_blank\" href=\"https://console.anyscale.com/template-preview/audio-dataset-curation-llm-judge\"><img src=\"https://img.shields.io/badge/\ud83d\ude80 Run_on-Anyscale-9hf\"></a>&nbsp;\n",
     "  <a href=\"https://github.com/anyscale/templates/tree/main/templates/audio-dataset-curation-llm-judge\" role=\"button\"><img src=\"https://img.shields.io/static/v1?label=&message=View%20On%20GitHub&color=586069&logo=github&labelColor=2f363d\"></a>&nbsp;\n",
     "</div>\n",
     "\n",
-    "**⏱️ Time to complete**: 30 min\n",
+    "**\u23f1\ufe0f Time to complete**: 30 min\n",
     "\n",
     "This tutorial demonstrates a batch inference pipeline that converts raw\n",
     "audio files into a curated subset using two different ML models.\n",
@@ -21,7 +21,7 @@
     "2. Resample each clip to 16 kHz for compatibility with Whisper.\n",
     "3. Transcribe the audio with the `openai/whisper-large-v3-turbo` model.\n",
     "4. Judge the educational quality of each transcription with a small Llama-3 model.\n",
-    "5. Persist only clips that score ≥ 3 to a Parquet dataset.\n",
+    "5. Persist only clips that score \u2265 3 to a Parquet dataset.\n",
     "\n",
     "Ray Data is particularly powerful for this use case because it:\n",
     "- **Parallelizes work** across a cluster of machines automatically\n",
@@ -259,7 +259,10 @@
     "        self.processor = AutoProcessor.from_pretrained(TRANSCRIPTION_MODEL)\n",
     "\n",
     "    def __call__(self, batch):\n",
-    "        token_ids = batch.pop(\"token_ids\")\n",
+    "        # Whisper's tokenizer probes token_ids with plain-Python truthiness\n",
+    "        # (`if not token_ids`, `if has_prompt`, ...), so hand it plain nested lists\n",
+    "        # rather than a pandas Series of numpy arrays.\n",
+    "        token_ids = [ids.tolist() for ids in batch.pop(\"token_ids\")]\n",
     "        transcription = self.processor.batch_decode(token_ids, skip_special_tokens=True)\n",
     "        batch[\"transcription\"] = transcription\n",
     "        return batch\n",
@@ -277,7 +280,7 @@
     "## LLM-based quality filter\n",
     "\n",
     "A Llama-3 model serves as a *machine judge* that scores each transcription\n",
-    "from 1 👎 to 5 👍 on its educational value. The **LLM Processor** API wraps the heavy\n",
+    "from 1 \ud83d\udc4e to 5 \ud83d\udc4d on its educational value. The **LLM Processor** API wraps the heavy\n",
     "lifting of batching, prompt formatting, and vLLM engine interaction using a declarative API style.\n",
     "\n",
     "Ray Data provides a high-level API for integrating LLMs into data pipelines. The preprocessing and postprocessing functions handle data preparation and result parsing."

diff --git a/templates/audio-dataset-curation-llm-judge/README.md b/templates/audio-dataset-curation-llm-judge/README.md
@@ -183,7 +183,10 @@ class Decoder:
         self.processor = AutoProcessor.from_pretrained(TRANSCRIPTION_MODEL)
 
     def __call__(self, batch):
-        token_ids = batch.pop("token_ids")
+        # Whisper's tokenizer probes token_ids with plain-Python truthiness
+        # (`if not token_ids`, `if has_prompt`, ...), so hand it plain nested lists
+        # rather than a pandas Series of numpy arrays.
+        token_ids = [ids.tolist() for ids in batch.pop("token_ids")]
         transcription = self.processor.batch_decode(token_ids, skip_special_tokens=True)
         batch["transcription"] = transcription
         return batch

diff --git a/templates/audio-dataset-curation-llm-judge/python_depset.lock b/templates/audio-dataset-curation-llm-judge/python_depset.lock
@@ -1,5 +1,5 @@
 --index-url https://pypi.org/simple
---extra-index-url https://download.pytorch.org/whl/cu128
+--extra-index-url https://download.pytorch.org/whl/cu130
 
 cffi==2.0.0 \
     --hash=sha256:00bdf7acc5f795150faa6957054fbbca2439db2f775ce831222b66f192f03beb \

diff --git a/templates/audio-dataset-curation-llm-judge/requirements.txt b/templates/audio-dataset-curation-llm-judge/requirements.txt
@@ -1,5 +1,5 @@
 soundfile==0.13.1
-# Match the base image's numpy (anyscale/ray-llm:2.55.1 ships numpy 1.26.4). The test
+# Match the base image's numpy (anyscale/ray-llm:2.56.0-py312-cu130 ships numpy 1.26.4). The test
 # layers this lock on the stock image with `uv pip install --no-deps`, so a higher numpy
 # would force-upgrade it and break the image's numpy-1.x-compiled pandas/scipy (ABI).
 numpy==1.26.4