leaderonehit/DRT-RL-8B-training-data

DRT RL Training Data

Paper: DRT: Dense Reasoning Trace for Efficient and Grounded Multimodal Reasoning

Original files from the DRT project's VisionR1_DAPO data directory, with parquet contents preserved. Machine-specific paths have been removed from metadata.

File	Rows	Bytes	Purpose
train.parquet	14,724	491,270,387	Original VisionR1 + DAPO training set
test.parquet	618	26,499,711	Evaluation split
mmfinereason_train.parquet	32,362	1,157,889,880	Additional mmfinereason training set
mmfinereason_train_summary.json	-	581	Preprocessing statistics (paths sanitized)

The mmfinereason set is an additional experimental dataset, not part of the default training split. Use the original train.parquet when reproducing the VisionR1 + DAPO run; do not automatically concatenate the two training files. The original train and test files also have separate configurations because extra_info.source_index exists in train but not test. Images are embedded as bytes in the parquet files.

Download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="leaderonehit/DRT-RL-8B-training-data",
    repo_type="dataset",
    local_dir="training_data/VisionR1_DAPO",
)

Dataset configurations can be loaded separately:

from datasets import load_dataset

train = load_dataset("leaderonehit/DRT-RL-8B-training-data", "visionr1_dapo_train", split="train")
test = load_dataset("leaderonehit/DRT-RL-8B-training-data", "visionr1_dapo_test", split="test")
extra = load_dataset("leaderonehit/DRT-RL-8B-training-data", "mmfinereason")

Provenance

• Code: https://github.com/HIT-leaderone/DRT
• RL checkpoint: https://huggingface.co/leaderonehit/DRT-RL-8B
• SFT data: https://huggingface.co/datasets/leaderonehit/DRT-SFT-8B-training-data

data_manifest.json records the original file sizes, row counts, schemas and SHA-256 checksums. Parquet contents and scientific statistics are preserved; machine-specific metadata is sanitized. This release does not assign a new license to the upstream datasets; consult their respective terms.