ahamad-ai/mineru-decoder-finetune-dataset

MinerU Decoder Fine-Tune Dataset (real data)

Note: the dataset viewer is disabled — the data is distributed as zip bundles, not a HuggingFace-loadable table. Download + unpack with the reproduce script (below), don't use load_dataset.

68,798 real image→tagged-text samples for teaching a VLM to emit inline formatting tags (bold, italic, sup, sub, underline, strike). Sources: 22 arXiv categories (bold/italic/sup/sub) + Washington & Texas legislative bills (underline/strike). Full write-up: DecoderFineTunedExperiment.md in the code repo.

Files

• FineTuneDataset_masters.zip — one_word / multi_word / multi_line masters (images + metadata + manifest)
• lora_adapter_combined_both.zip — trained LoRA adapter for opendatalab/MinerU2.5-2509-1.2B

Use

# from the code repo:
python scripts/reproduce_finetune_dataset.py            # downloads + unzips + regenerates splits
python scripts/reproduce_finetune_dataset.py --with-adapter

Or manually:

from huggingface_hub import hf_hub_download
z = hf_hub_download("ahamad-ai/mineru-decoder-finetune-dataset", "FineTuneDataset_masters.zip", repo_type="dataset")