README.md

metadata

library_name: datasets
tags:
  - turkish
  - english
  - azerbaijani
  - turkmen

Merged Datasets 11-12

This dataset is a compilation of various sources filtered for length and quality.

Statistics

Dataset	Language	Count	Mean Len	Median Len	Max Len	Min Len
Ba2han/synth-tr	Turkish	33,416	2874	2692	17714	400
Ba2han/synth-2m-v2	Turkish	2,065,969	2894	2718	110664	358
Ba2han/filtered-cosmos	Turkish	785,340	4987	4646	11000	1000
Ba2han/fineweb2-filtered-tr	Turkish	5,896,067	5853	5265	15000	304
Ba2han/finePDF-filtered-tr	Turkish	124,223	8294	8273	15000	301
Ba2han/vngrs-web-filtered	Turkish	2,805,334	5640	5115	15000	300
Ba2han/HPLT2-filtered-edu-tr	Turkish	258,132	5835	5001	15000	501
Ba2han/c4_tr_fineweb-filtered	Turkish	374,466	5841	4989	15000	316
openbmb/Ultra-FineWeb	English	2,500,000	2085	1729	5800	250
LocalDoc/community_oscar_azerbaijani	Azerbaijani	2,000,000	2025	1699	5800	250
LocalDoc/azerbaijani-english-parallel-corpus	Azerbaijani-English	759,133	354	330	824	250
LocalDoc/news_azerbaijan_2	Azerbaijani	697,188	1432	898	5800	250
ituperceptron/turkish-math-186k	Turkish	179,417	1375	1097	5800	250
ryzentm/turkmence_all_data	Turkmen	33,327	998	803	3948	250
BILGEM-AI/BILGE-Wiki-Tr-Plus	Turkish	469,967	1855	1101	5800	250

Processing Details

1. Batch 1: Dropped examples < 300 characters.
2. Batch 2: Dropped examples < 250 characters and > 5800 characters.
3. Special Handling:
   • BILGE-Wiki: Split long articles by .\n, chunks aimed at ~5500 chars.
   • Ultra-FineWeb: Filtered score > 0.54, shuffled, capped at 2.5M.