Missing “wemath (WeMath data)” zip file in LongVT-Source dataset training data
Hello,
I am using the LongVT-Source dataset for my project. In the “Training Data” table of the dataset card, there is a row for “wemath — WeMath data — 1 zip”. However, when I downloaded the dataset, I was unable to find the corresponding zip file for this WeMath data.
Could you please check whether the file was uploaded properly or there was some error?
If this “wemath” zip is missing, could you kindly upload it (or guide me how to obtain it)? If it's present but I missed it, could you indicate its correct path / filename?
Thank you very much for maintaining this valuable dataset. I appreciate any clarification or help.
Sorry about that — we indeed missed uploading the WeMath data. We have now uploaded all of its source images to wemath_1.zip.
Thank you so much for your attention. If you notice anything else we may have missed, please feel free to let us know!
Hi again,
Thanks for uploading wemath_1.zip and letting me know. I appreciate your fast attention.
However, after updating to include the WeMath data, I encountered two further issues when trying to run SFT training with the full dataset (source + annotations). I want to check if these are known issues, or if perhaps some data/metadata is inconsistent.
1. Annotation alignment error when loading LongVT-Parquet data
When loading the annotation data from LongVT-Parquet, the loading pipeline raises a ValueError indicating that the “features can't be aligned” — specifically complaining about mismatched types for the messages field (expected a structure containing image_url with "url": string, but instead got records with image_url set to null). This suggests that, at least for some samples, the annotation expects image/video content but the corresponding URLs or media references don't match the expected schema.
Full error (abridged):
Error executing job with overrides: ['config_yaml=/home/LongVT/examples/sft/longvt_7b_sft.yaml']
Traceback (most recent call last):
File "/mnt/SSD2_16TB/LongVT/lmms-engine/src/lmms_engine/launch/cli.py", line 112, in main
task.build()
File "/mnt/SSD2_16TB/LongVT/lmms-engine/src/lmms_engine/train/runner.py", line 59, in build
self.train_dataset = self._build_train_dataset()
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/SSD2_16TB/LongVT/lmms-engine/src/lmms_engine/train/runner.py", line 118, in _build_train_dataset
dataset.build()
File "/mnt/SSD2_16TB/LongVT/lmms-engine/src/lmms_engine/datasets/iterable/base_iterable_dataset.py", line 49, in build
self._build_from_config()
File "/mnt/SSD2_16TB/LongVT/lmms-engine/src/lmms_engine/datasets/iterable/multimodal_iterable_dataset.py", line 92, in _build_from_config
self.data_list, self.data_folder = DataUtilities.load_inline_datasets(self.config.datasets)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/SSD2_16TB/LongVT/lmms-engine/src/lmms_engine/utils/data_utils.py", line 225, in load_inline_datasets
data_list = concatenate_datasets(data_list)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/miniconda3/envs/lmms/lib/python3.11/site-packages/datasets/combine.py", line 213, in concatenate_datasets
return _concatenate_map_style_datasets(dsets, info=info, split=split, axis=axis)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/miniconda3/envs/lmms/lib/python3.11/site-packages/datasets/arrow_dataset.py", line 6485, in _concatenate_map_style_datasets
_check_if_features_can_be_aligned([dset.features for dset in dsets])
File "/home/miniconda3/envs/lmms/lib/python3.11/site-packages/datasets/features/features.py", line 2305, in _check_if_features_can_be_aligned
raise ValueError(
ValueError: The features can't be aligned because the key messages of features {'id': Value('string'), 'messages': List({'content': List({'image_url': Value('null'), 'text': Value('string'), 'type': Value('string'), 'vi
deo_url': {'url': Value('string')}}), 'role': Value('string')})} has unexpected type - List({'content': List({'image_url': Value('null'), 'text': Value('string'), 'type': Value('string'), 'video_url': {'url': Value('str
ing')}}), 'role': Value('string')}) (expected either List({'content': List({'image_url': {'url': Value('string')}, 'text': Value('string'), 'type': Value('string'), 'video_url': {'url': Value('string')}}), 'role': Value
('string')}) or Value("null").
Because of this, the dataset fails to load/concatenate properly, preventing me from proceeding to training.
Could you please check whether the Parquet annotation schema has been updated recently, and if so, whether all entries conform to the expected schema (i.e. messages with valid image_url / video_url)?
2. Missing source media files even after download and unzip
After I bypassed the first issue (editing the code to skip/allow null image_url), another problem arose in actual iteration through the dataset. For some samples (e.g. from videor1), when the code tries to open image files (or video frames) listed in the annotation, it fails with FileNotFoundError.
For example:
Traceback (most recent call last):
File "/mnt/SSD2_16TB/LongVT/lmms-engine/src/lmms_engine/datasets/iterable/multimodal_iterable_dataset.py", line 182, in __iter__
data_dict = self.get_one_sample(self.cur_idx, curr_data_folder[self.cur_idx], curr_data_list)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/SSD2_16TB/LongVT/lmms-engine/src/lmms_engine/datasets/iterable/multimodal_iterable_dataset.py", line 127, in get_one_sample
data_dict = self.load_from_json(data_list[index], data_folder)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/SSD2_16TB/LongVT/lmms-engine/src/lmms_engine/datasets/iterable/vision_iterable_dataset.py", line 83, in load_from_json
images = [Image.open(os.path.join(data_folder, image)) for image in images_list]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/SSD2_16TB/LongVT/lmms-engine/src/lmms_engine/datasets/iterable/vision_iterable_dataset.py", line 83, in <listcomp>
images = [Image.open(os.path.join(data_folder, image)) for image in images_list]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/miniconda3/envs/lmms/lib/python3.11/site-packages/PIL/Image.py", line 3493, in open
fp = builtins.open(filename, "rb")
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
FileNotFoundError: [Errno 2] No such file or directory: '/home/LongVT/data/source/data/videor1/Math/Multimath-300k/50128c41d4d4d905bfd110504eca9671.png'
[2025-12-05 19:15:26,272][qwen_vl_utils.vision_process][INFO] - decord: video_path='file:///home/LongVT/data/source/data/videor1/CLEVRER/train_videos/video_09643.mp4', total_frames=128, video_fps=25.0,
time=0.019s
Traceback (most recent call last):
File "/mnt/SSD2_16TB/LongVT/lmms-engine/src/lmms_engine/datasets/iterable/multimodal_iterable_dataset.py", line 182, in __iter__
data_dict = self.get_one_sample(self.cur_idx, curr_data_folder[self.cur_idx], curr_data_list)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/SSD2_16TB/LongVT/lmms-engine/src/lmms_engine/datasets/iterable/multimodal_iterable_dataset.py", line 127, in get_one_sample
data_dict = self.load_from_json(data_list[index], data_folder)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/SSD2_16TB/LongVT/lmms-engine/src/lmms_engine/datasets/iterable/vision_iterable_dataset.py", line 83, in load_from_json
images = [Image.open(os.path.join(data_folder, image)) for image in images_list]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/SSD2_16TB/LongVT/lmms-engine/src/lmms_engine/datasets/iterable/vision_iterable_dataset.py", line 83, in <listcomp>
images = [Image.open(os.path.join(data_folder, image)) for image in images_list]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/miniconda3/envs/lmms/lib/python3.11/site-packages/PIL/Image.py", line 3493, in open
fp = builtins.open(filename, "rb")
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
FileNotFoundError: [Errno 2] No such file or directory: '/home/LongVT/data/source/data/videor1/Chart/MapQA/images/map_3585.png'
This suggests that although the source dataset shows "videor1" (and other categories) zips, some of the media files referenced in the annotations are not present in the unpacked source folders.
Could you please help check:
Whether all training-source zips were re-uploaded (or extracted) properly, especially for categories like videor1, wemath, longvideoreason, etc.
Whether there might have been some files omitted, corrupted, or path mismatches (e.g. wrong directory naming / subfolder structure) between the annotation JSON/Parquet and the actual unpacked media.
Thank you again for your support and hard work on this dataset and framework. I look forward to your reply.
Thanks a lot for reporting this issue and for taking the time to try out the dataset.
Before we deep-dive into the bug on our side, could I first check with you: have you already replaced the paths in the parquets (e.g., image_url / video_url) and the paths that appear inside some system prompts with your own local absolute paths?
We intentionally sanitized/cleaned these fields before releasing the annotations (parquets) to avoid any potential privacy leakage, so the original paths are not directly usable and need to be customized by each user.
Thank you for the explanation. Yes — I have already replaced the paths in all Parquet files by configuring data_folder in my training YAML (e.g., videor1, longvideoreason, wemath, etc.). The loader concatenates data_folder with the relative paths in the annotation, and this part works correctly — I can confirm that videos are successfully loaded, for example:
[2025-12-05 19:15:26,272][qwen_vl_utils.vision_process][INFO] - decord: video_path='file:///home/LongVT/data/source/data/videor1/CLEVRER/train_videos/video_09643.mp4', total_frames=128, video_fps=25.0,
time=0.019s
So the local path mapping mechanism is functioning as expected.
However, the errors I’m encountering are different. Some samples reference image files that do not exist anywhere in the downloaded source data, even after full extraction. For example:
FileNotFoundError: [Errno 2] No such file or directory: '/home/LongVT/data/source/data/videor1/Math/Multimath-300k/50128c41d4d4d905bfd110504eca9671.png'
FileNotFoundError: [Errno 2] No such file or directory: '/home/LongVT/data/source/data/videor1/Chart/MapQA/images/map_3585.png'
I manually searched the entire extracted dataset, and these files (50128c41d4d4d905bfd110504eca9671.png, map_3585.png) simply do not exist in any zip.
This suggests that some sanitized paths in the Parquet refer to media that is not included in the released source data, rather than an issue with my local path replacement.
Given that some samples load correctly, it seems that only a subset of entries in the Parquet reference missing assets.
Could you confirm whether these missing image files were supposed to be included in the videor1 source data?
If not, should users filter out such samples, or will the full media set be re-uploaded?
This might be relevant to lmms-engine loading consistency.
Thanks again for your help and for maintaining this large dataset!
Thanks a lot for bringing this up! After a rigorous cross-check between my local media files and the remote HF repo, I can confirm that I indeed missed uploading seven zip files in the Video-R1 series (videor1_7.zip to videor1_13.zip). All of them have been uploaded.
My sincere apologies for any troubles caused due to my carelessness. If you notice any other issues, please do not hesitate to let me know.