# Aadya

> Aadya (aadya-m1 and aadya-m1-mini) is the best open-source model for forecasting the next period: first by mean error on all three public benchmarks (AadyaBench). It gives a probability for every possible cycle length, not a single date, and runs on the user's device so cycle data never leaves it. Not a medical device; no diagnosis, contraception or fertility claims.

Key facts, each traceable to a committed benchmark report:
- Standing: best accuracy among open-source models on AadyaBench. Lowest mean error of every open method compared (Bayes reference, same-data LSTM, Poisson-with-skips, rolling median, rolling mean, exponential smoothing, 28-day) on all three public boards (539 simulated users; Creighton, 251 real users; Marquette, 113 real users). Scope: the real-data lead over the Bayes reference is about 1% of the error and within noise on Marquette alone; on an unseen simulator it is slightly behind the Bayes reference and the LSTM; mid-cycle without the ovulation-test input a Poisson-with-skips model is slightly ahead (2.34 vs 2.43).
- aadya-m1 has 256M parameters; aadya-m1-mini has 0.84M parameters (3.4 MB). Same inputs and outputs. Weights are Apache-2.0 on Hugging Face.
- Input: past cycle lengths (days), days since the last period started, optional age group, optional ovulation-test day. Output: a probability for each cycle length from 1 to 120 days.
- Error is measured with CRPS (lower is better). aadya-m1: 2.92 vs 3.70 for the usual tracker method on 539 simulated users (21% lower); 1.93 vs 2.35 on 364 public real users (18% lower).
- Against the strongest research baseline (a hierarchical Bayes reference) the lead on real users is small, about 1% of the error. With no history at all it is level with a plain 28-day baseline (2.18 vs 2.13).
- An optional, experimental ovulation-test input narrowed the 80% window from about 12 to about 7 days in one study of 41 people (1.70 vs 3.49 for the usual method). It is never presented as an ovulation date.
- Known limits: very irregular cycles, postpartum and perimenopause phases, hormonal contraception and very short histories are harder; forgotten period logs hurt every method; evidence is simulated users plus two public cohorts, not a clinical trial.

## Pages
- [Home](https://aadya.manasdutta.com/): what Aadya is and the headline results
- [Models](https://aadya.manasdutta.com/models/): aadya-m1 and aadya-m1-mini, specs, which to use
- [aadya-m1](https://aadya.manasdutta.com/models/m1/): the 256M model: specs, accuracy, inputs and outputs, limits
- [aadya-m1-mini](https://aadya.manasdutta.com/models/m1-mini/): the 0.84M model for phones: specs, accuracy against aadya-m1, limits
- [Benchmarks](https://aadya.manasdutta.com/benchmarks/): AadyaBench leaderboard, paired confidence intervals, real-data tracks
- [How it works](https://aadya.manasdutta.com/how-it-works/): from logged cycles to a calibrated distribution; interactive examples
- [Case studies](https://aadya.manasdutta.com/case-studies/): wins, ties and losses, including negative results
- [Privacy and safety](https://aadya.manasdutta.com/privacy/): no backend, no accounts, no analytics; what Aadya is not
- [Try it](https://aadya.manasdutta.com/try/): real model output you can move around
- [Docs](https://aadya.manasdutta.com/docs/): quickstart, inputs and outputs, integration rules, FAQ
- [Research](https://aadya.manasdutta.com/research/): paper, citation, evaluation data

## Links
- [Model on Hugging Face](https://huggingface.co/manasdutta04/aadya-m1)
- [Mini model on Hugging Face](https://huggingface.co/manasdutta04/aadya-m1-mini)
- [Paper (Zenodo preprint)](https://doi.org/10.5281/zenodo.23236800)
- [GitHub](https://github.com/manasdutta04/aadya)
- [Full text for LLMs](https://aadya.manasdutta.com/llms-full.txt)

## Optional
- [Sitemap](https://aadya.manasdutta.com/sitemap.xml)