# chartnet_qa_06_difficulty Download the original Parquet files for the `core_permissive` configuration of [ibm-granite/ChartNet](https://huggingface.co/datasets/ibm-granite/ChartNet/tree/ff6cf062b8bb28c9c29ca315f18b1b890ac39712), revision `ff6cf062b8bb28c9c29ca315f18b1b890ac39712`. Images are embedded in `image`; save each row's original bytes as `.png` at the release root. Numeric row positions are not image IDs. Map each JSONL image filename to the source Parquet `id` column: `.png` is written from that row's `image["bytes"]`. The release record's top-level `id` is a conversation identifier, so use its image filename for the lookup. Run this example from the release root after installing `pyarrow` and downloading the source shards into `/downloads/source/`: ~~~python from pathlib import Path import json import pyarrow.parquet as pq dataset = "chartnet_qa_06_difficulty" needed = set() with (Path(dataset) / (dataset + ".jsonl")).open() as f: for line in f: for message in json.loads(line)["messages"]: for item in message["content"]: if isinstance(item, dict) and item.get("type") == "image": needed.add(item["image"]) produced = set() for path in sorted(Path("/downloads/source").rglob("*.parquet")): for batch in pq.ParquetFile(path).iter_batches(columns=["id", "image"]): for row in batch.to_pylist(): filename = str(row["id"]) + ".png" if filename in needed: data = row["image"]["bytes"] if data is None: raise ValueError("Missing embedded image: " + filename) Path(filename).write_bytes(data) produced.add(filename) missing = sorted(needed - produced) if missing: raise FileNotFoundError(f"{len(missing)} requested source images were not produced: {missing[:20]}") print(f"Produced all {len(needed)} requested images") ~~~