Aruna is a vision-language model trained on 220+ classical Jyotish texts. Show it a birth chart image — it reads the planets, houses, and nakshatras, and writes an interpretation grounded in real astrological literature.
Ten training runs, a three-modality data engine, and a from-scratch astronomical verification system. This is not a concept — it's a model that's already learning. The engine feeds Aruna three complementary kinds of training data at once — so it learns to see charts, speak the tradition, and respect the astronomy in the same run.
29,954 quality-filtered pairs of real birth-chart images matched with interpretation text, extracted from 220+ Jyotish books — this teaches Aruna to see a chart.
2,000 OCR-cleaned passages drawn from the tradition's source literature — this teaches Aruna to speak in the language and reasoning style of practitioners.
5,000 chart–interpretation pairs generated by the in-house Swiss Ephemeris engine — every planetary position astronomically exact, so Aruna learns from data that is provably correct.
Nine completed runs, each building on the last: loss fell from ~1.42 at baseline to 1.18 at Run #9 — with the model reproducing dense KP nakshatra tables with high fidelity by the final checkpoint. Run #10 is training now, the first with chart images, book text, and verified chart data learning together.
220+ classical and modern Jyotish texts extracted into 34,318 image–interpretation pairs, quality-gated to 29,954 clean examples, plus an OCR-cleaned text corpus and a live web-scraping pipeline still adding sources. Every file screened against strict quality rules before entering training.
A from-scratch calculation engine built on the Swiss Ephemeris — the same astronomical data professional Jyotish software uses — with a 2,063-city atlas. It generated 5,000 deterministic chart–interpretation pairs, so Aruna's readings can be checked against real planetary positions.
The path ahead is mapped: a held-out evaluation benchmark, targeted gap-fix runs, a final consolidation run, and a public release — targeted for October 2026. The architecture docs, production plan, and model card are already written.
No chart-reading expertise required — Aruna does the technical reading for you.
Any standard North- or South-Indian style birth chart. A photo, a screenshot, an export from chart software — Aruna reads the layout directly.
Planets, houses, and nakshatras are identified from the image itself — the same way a trained practitioner would scan the chart.
Aruna writes an interpretation shaped by the classical texts it was trained on — not generic, AI-flavored astrology copy.
220+ classical and modern Jyotish texts — the actual books practitioners study from, not summaries of summaries — plus an 86-book text corpus feeding the language of the tradition directly into training.
A companion astronomical engine calculates real planetary positions using the same ephemeris data professional Jyotish software relies on.
Aruna reads the chart image itself — it doesn't need positions typed in manually to produce an interpretation.
A QLoRA fine-tune of Qwen2.5-VL-3B-Instruct, trained across three data modalities
Aruna is a pre-release research project moving toward a production build.
Next milestone: a held-out evaluation benchmark scoring planet/house placement accuracy, nakshatra accuracy, and interpretation quality — the last gate before a production release.