A repository on GitHub's trending page is not usually a story about data volume. Renhuai123/ziwei-doushu is worth the second look because the numbers are strange: 674,976 horoscopes times 13 analysis dimensions gives 8,774,688 interpretation passages, roughly 30,000 characters per chart, about 20.4 billion characters of prose in total. Uncompressed that is 56.6 gigabytes; compressed with zstd across 77 per-year files it is still 161 megabytes. Released on 15 October 2026 under CC BY 4.0 -- commercial use allowed, attribution required. The code is MIT.
__The method is the point__
None of it was written by a language model. The corpus is generated deterministically by a rules engine plus a hand-curated library of statements: same input, same output, every time. The project is explicit that the generation path does not pass through an AI. The engine handles true-solar-time correction by longitude, China's 1986-1991 daylight saving, four separate conventions for leap months, the late-zi hour day boundary, seven brightness grades per star and more than 1,100 lines of pattern rules.
That is the interesting part: not a model inventing text, but a rule system wrestled out of a domain and then turned into a training corpus. The repository lists its uses as fine-tuning small models, retrieval for RAG, and A/B baselines whenever the rule library changes.
__The critical read__
Twenty billion characters is not twenty billion characters of information. A deterministic generator recombining a fixed statement library produces enormous volume with bounded diversity. Train a model on it and you may mainly teach it the template. For retrieval, for evaluation, and as a reference corpus this is excellent; as pretraining data its value is limited by the library, not by the byte count.
The project is unusually candid elsewhere too. It documents that the older v3.0 sample set contains 2,520 entries with impossible dates -- 30 February and the like -- because the generator took days 1 to 30 of every month, and that the open-source interface ships a placeholder where the interpretations should be. The commercial site that funds the work stays closed.
__Why it matters__
The largest text corpora no longer come only from scraping or from a model emitting text. Here a single, part-time website for Chinese astrology delivered 56.6 gigabytes of it from a hand-written rules engine, free for commercial use -- and 1,126 people starred the repository in a single day.
That is not an announcement but a data point: the cost of producing domain-sized text has fallen far enough that a niche project does it on the side. The question that follows is uncomfortable and reaches well beyond astrology. When volume is cheap, scarcity -- real observation, real variation, real surprise -- becomes the only thing still worth paying for.
