Train Dataset Source

[sysuwzy]

Thanks for your fantastic work!

Could you please tell me your train data for MOSS-VL include which open-source datasets?

[xmqian03]

Hi! Thanks for your interest in MOSS-VL.

Our training data consists of (1) open-source datasets collected from the community and (2) in-house synthetic data. We have compiled the list of the 205 open-source datasets used in our training here: docs/open_source_datasets.md

Note that these datasets went through our internal curation pipeline (including format unification, quality filtering, deduplication, and additional cleaning and re-annotation) before use, so the actual training samples may differ from the raw public releases.

Hope this helps — feel free to ask if you have any further questions!

[sysuwzy]

Thanks for your kind response!

[sysuwzy]

Thanks for your datasets list! May I ask about the usage of their splits? When training MOSS-VL, was only the train split used, no test or validation splits?

[xmqian03]

Hi, thanks for the follow-up!

No test or validation split was used — we only train on the training splits of these datasets. On top of that, our curation pipeline performs strict decontamination against the test/validation sets of all our evaluation benchmarks, and any overlapping sample is removed before training. Datasets without official splits go through the same decontamination step.

Decontaminating against test sets is the most basic rule of any serious training pipeline — after all, nobody trains on the test set :)

Hope this clarifies it!