Error code:   ConfigNamesError
Exception:    ValueError
Message:      Some splits are duplicated in data_files: ['train', 'train', 'train', 'train', 'train', 'train', 'train', 'train', 'validation']
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/dataset/config_names.py", line 67, in compute_config_names_response
                  config_names = get_dataset_config_names(
                      path=dataset,
                      token=hf_token,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 161, in get_dataset_config_names
                  dataset_module = dataset_module_factory(
                      path,
                  ...<4 lines>...
                      **download_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1217, in dataset_module_factory
                  raise e1 from None
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1192, in dataset_module_factory
                  ).get_module()
                    ~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 646, in get_module
                  patterns = sanitize_patterns(next(iter(metadata_configs.values()))["data_files"])
                File "/usr/local/lib/python3.14/site-packages/datasets/data_files.py", line 151, in sanitize_patterns
                  raise ValueError(f"Some splits are duplicated in data_files: {splits}")
              ValueError: Some splits are duplicated in data_files: ['train', 'train', 'train', 'train', 'train', 'train', 'train', 'train', 'validation']

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

OpenMMReasoner-RL-74K — RL Source Mix

74,971 train rows · 1,724 val rows · 8 train sources + 1 val source · images embedded

This is the source data mix for the OpenMMReasoner (OMR) RL stage, distributed as per-source parquet files. It was used to construct the omr-grpo-train training parquet. Each file corresponds to one upstream benchmark/dataset that was filtered, reformatted into verl-compatible schema, and had images embedded as bytes.

This dataset is derived from the OpenMMReasoner project.

Train Files — Per-Source Row Counts

File	`data_source`	Rows	Description
`algopuzzle_train.parquet`	`algopuzzle`	1,800	AlgoPuzzle: algorithmic visual puzzle reasoning
`mmk12_train.parquet`	`mmk12`	4,701	MM-K12: multimodal K-12 math problems
`puzzlevqa_train.parquet`	`puzzlevqa`	2,000	PuzzleVQA: visual puzzle QA
`thinklite_vl_hard_train.parquet`	`thinklite_vl_hard`	10,960	ThinkLite-VL hard split: challenging visual math
`tqa_train.parquet`	`tqa`	6,501	TQA: textbook-figure question answering
`virl39k_train.parquet`	`virl39k`	38,870	VIRL-39K: diverse visual instruction reasoning
`wemath_pro.parquet`	`wemath_pro`	4,296	WeMath professional/competition-level math
`wemath_standard.parquet`	`wemath_standard`	5,843	WeMath standard difficulty math
Train total		74,971

Val File — Per-Source Row Counts

File	`data_source`	Rows
`val.parquet`	`lmms_eval_aime_2024`	240
	`lmms_eval_aime_2025`	240
	`lmms_eval_amc23`	40
	`lmms_eval_mathvision_reason_testmini`	304
	`lmms_eval_mmmu_val_thinking`	900
Val total		1,724

Schema

Each train parquet shares the core verl-compatible schema (exact columns vary slightly per source due to upstream metadata columns retained):

Column	Type	Description
`prompt`	`list<struct<role: string, content: string>>`	Chat messages: system (CoT instruction) + user (question with `<image>` token)
`images`	`list<struct<bytes: binary, path: string>>`	Embedded image(s) as PNG bytes
`data_source`	`string`	Source identifier
`ability`	`string`	Reasoning ability tag
`reward_model`	`struct<ground_truth: string, style: string>`	Ground-truth answer + reward style
`tokens`	`int64`	Pre-computed token count estimate
`avg_reward`	`double`	Average reward from the OMR RL run (populated post-training)

Additional source-specific metadata columns (e.g. options, solution, id, subject, category, qid) are present in individual files.

The val.parquet uses extra_info: struct<answer, index, question, split> instead of reward_model.

Relationship to `omr-grpo-train`

The omr-grpo-train dataset is built from these source files by:

Merging all 8 train parquets
Filtering rows whose tokenized prompt exceeds max_prompt_length under Qwen3-VL
Patching the system prompt to the OMR-style CoT instruction uniformly
Re-encoding as a single consolidated parquet

The resulting omr_full_v1.parquet (74,971 rows) has identical source distribution to these files.

Source Projects

This dataset is derived from:

OpenMMReasoner: https://huggingface.co/OpenMMReasoner
VIRL-39K: https://huggingface.co/datasets/VIRL-Bench/VIRL-39K
ThinkLite-VL: https://huggingface.co/datasets/VLM-Reasoner/ThinkLite-VL
TQA: https://huggingface.co/datasets/derek-thomas/ScienceQA
WeMath: https://huggingface.co/datasets/We-Math/We-Math
MM-K12: https://huggingface.co/datasets/Zhiqiang007/MathV360K
PuzzleVQA: https://huggingface.co/datasets/declare-lab/PuzzleVQA
AlgoPuzzle: https://huggingface.co/datasets/declare-lab/AlgoVQA

Citation

@misc{openmmreasoner2025,
  title  = {OpenMMReasoner: Open-Source Multimodal Math Reasoner via Reinforcement Learning},
  year   = {2025},
  url    = {https://huggingface.co/OpenMMReasoner}
}

License

Each file inherits the license of its upstream source dataset. Users must comply with the terms of the respective source datasets listed above.

Downloads last month: 19

Total file size:

6.54 GB