Overtokenized sentences
Love the fact that this exists, thanks! I have been working on training a Sindhi transformer, I'll see how well it does when adding this dataset.
I find that there are some sentences which are overtokenized, such as
ه ڪ م ح ب ت ي ن ي ا ڻ ي َء پ ا ر ا ن ه ڪ ع ظ ي م و ا ل د ج ي ج يَ و ن ڪ ٿ ا ج ي ه ڪ ج ه ل ڪ
ج ي ن ا ل ي س ا ن س ڃ ا ت و و ڃ ي ٿ و پ ش ا ور، ملتان ۽ دهليَء ۾ رهڻ کان پوِء ع ۾ هجرت
پ ي ا ر ڪ ر ڻ س ک و ز ب ي د ه م ص ط ف ي ا ن ي ت ا ڪ م ل ي شَ و ر
Seems to be about 4% of the corpus
Hi, thanks for the kind words and for catching this! You're exactly right about the data, it has a character-spacing artifact from the extraction phase.
Since programmatically fixing the word boundaries introduces linguistic noise, I'm deploying a density filter in the next release to preserve data purity.
Good luck with your transformer training! If you'd be open to collaborating on refining the preprocessing script or testing the next release together, I'd love to team up. Let me know!
That sounds great! I'm including the dataset with a collection I've built elsewhere when training, and I'd be happy to try new versions of the dataset.
How much of this was from Oscar Common Crawl or the community version available here on HF? Just wondering, since I collected that as well and wouldn't want too much duplication
Thanks for including it in your collection!
To answer your question: roughly 70% of the raw volume draws from existing open-source Sindhi datasets (including OSCAR, Common Crawl, and community HF uploads), while about 30% is my own collection I’ve been running since January.
However, you don't need to worry about duplication if you merge this with your current data. Because the entire 57M raw rows went through an aggressive exact-deduplication pipeline using cryptographic hashing, the overlap across all those platforms was completely stripped out.
By using this corpus, you're effectively getting a much cleaner, deduplicated version of the public OSCAR/HF data, plus the 30% net-new text.