t2i-150k

# UniReason T2I Reason (HuggingFace format) Converted from `Alex11556666/Reason_Tuning` original parquet+JSON layout into HuggingFace-`datasets`-compatible parquet shards. ## Stats - Rows: 148,471 - Shards: 297 - Missing (JSON entries whose image was not in the parquet shards): 0 ## Schema | Column | Type | Description | |---------------|-----------|------------------------------------------| | id | int64 | Sample id from original JSON | | prompt | string | Text prompt (originally `txt`) | | explanation | string | Short factual explanation | | reasoning | string | Long chain-of-thought reasoning | | image | Image | Target image rendered by Qwen-Image | | width | int64 | Image width | | height | int64 | Image height | | image_format | string | e.g. "PNG" | ## Usage ```python from datasets import load_dataset ds = load_dataset("parquet", data_dir="data", split="train") print(ds[0]["prompt"]) ds[0]["image"].save("first.png") ``` The `image` column is already in HuggingFace's Image feature format (`{"bytes": ..., "path": ...}`); `datasets` will lazy-decode to PIL. ## Source - Original dataset: [Alex11556666/Reason_Tuning](https://huggingface.co/datasets/Alex11556666/Reason_Tuning) - Paper: UniReason 1.0 (arXiv:2602.02437) - This subset corresponds to the **T2I Reasoning** part of Phase I in the paper.