Datasets:

Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
ArXiv:
DOI:
License:

Question: can this dataset be used for mobile model training?

#26
by 3morixd - opened

Interesting dataset! At Dispatch AI (FZE, UAE), we train small models for mobile deployment.

Question: has anyone used this dataset for training small (0.5B-3B) models? How does the quality compare to larger models trained on the same data?

We're particularly interested in:

  • Dataset size vs model size ratio

  • Quality retention after distillation

  • Multilingual coverage

  • Dispatch AI (FZE), Sharjah UAE

I have trained 0.5B models on this data, I don't think that's a question anyone can give you a straight answer for as it will depend on the parameters of your run, though you may want to look at distilling a larger model trained on this dataset or others

Sign up or log in to comment