Datasets:
[issue] incorrect data card, missing weborganizer labels
Howdy Kyle/Luca; There is a minor documentation mismatch on the dataset card for 150b mix. Document counts (~10B total) match the full Dolma3 pool; using Table 4 of the OLMo 3 technical report I realized the 150B mix actually uses ~104M documents total. When you have time I would update the data card table in the documentation. Thanks again for dolma/olmo 😊
| Source | Type | Tokens | Documents |
|---|---|---|---|
| Common Crawl | Web pages | 121B (76.9%) | 84.5M |
| olmOCR Science PDFs | Academic documents | 19.9B (12.6%) | 2.25M |
| Stack-Edu (Rebalanced) | GitHub code | 11.1B (7.06%) | 14.3M |
| arXiv | Papers with LaTeX | 1.29B (0.82%) | 247K |
| FineMath 3+ | Math web pages | 4.10B (2.60%) | 2.57M |
| Wikipedia & Wikibooks | Encyclopedic | 64.6M (0.04%) | 119K |
| Total | 157B (100%) | 104M |
Ps: My explatory analysis found that something like 15% of the shards outside of the CommonCrawl did not have the weborganizer topic domain, and none have the format domain. Assuming here that topic domain for missing shards is due to their provenance being obviously science/tech or something similar, and that it was an oversight to leave off the format domains. I have already enriched the data myself using the weborganizer classifier for the topics/formats, so let me know if a PR with that data would be helpful if desired.
Missing topic labels in <category (docs)>: stack_edu-Python (2.79M), stack_edu-Java (2.56M), finemath-3plus (2.27M), stack_edu-Cpp (1.63M), stack_edu-JavaScript (1.43M), stack_edu-Markdown (1.25M), stack_edu-PHP (1.15M), stack_edu-CSharp (1.06M), stack_edu-C (0.67M), stack_edu-Shell (0.51M)...
Hi @glennmatlin - thanks so much for checking out our work and for pointing out the dataset card typo. I just merged the fix so the card should be correct now!
Regarding the weborganizer topics -- this topic classifier is actually only applied for Common Crawl and olmOCR science pdfs, which is why you don't see it in the other sources.
Got it, thanks for the heads-up, @baileyk — if AI2 makes an "official" topic classification for those source groups, or the "official" format domains for dolma, please let me know! Otherwise, I will assume that stock weborganizer is good-to-go and run with whatever it spits out 😊 thanks again