Pick an example history and a day. Every curve here is the real aadya-m1-mini output, computed offline on made-up example data. Nothing you do on this page is sent anywhere.
When next onset is likely
Chance by day
P(onset within N days)
How much error it removes
Usual tracker method = rolling median of the last 3 cycles. Blue bars are also better than the strongest comparison we tested; grey bars are ties or not tested. Simulated tracks are marked "simulated". From the committed benchmark reports.
Simulated cases
364 public real users
Lower is better. aadya-m1-mini has the lowest mean error among the methods we ran on both cohorts, level with the full-size aadya-m1. Its lead over the Bayes reference is small (about 1% of the error): clear on Creighton, within noise on Marquette alone.
Which model should I use?
- aadya-m1-mini (this demo): 0.84M parameters, 3.4 MB. Phones, browsers, low-power devices.
- aadya-m1: 256M parameters, about 1 GB. More capacity for servers, desktops, research and fine-tuning.
- On the 364 public real users we have, the two score the same, so the small one costs nothing in accuracy there.
AadyaBench board
| # | Model | CRPS | vs hierarchical |
|---|
Curves are precomputed real model outputs on example inputs. Benchmark numbers are copied from the committed AadyaBench v2 report (simulated users; the real-data lines are small public cohorts). To try your own history, run the model on your own device: see the model card. Do not type real health data into websites. Not a medical device.
A few lines, on your own machine
from huggingface_hub import snapshot_download
import sys
path = snapshot_download("manasdutta04/aadya-m1-mini"); sys.path.insert(0, path)
from aadya_m1 import AadyaM1
model = AadyaM1.from_pretrained(path)
pmf = model.forecast([28, 30, 27, 29], days_since_last_period=0)
print(model.window(pmf)) # 80% window in cycle daysNeeds numpy, scipy, torch, safetensors and huggingface_hub. Your data stays on your device.
Safety and honesty
- Not a medical device. No diagnosis, no contraception or fertility advice, no "safe days".
- Probabilities, not promises. Show the whole window, never one date as certain.
- Evidence is early. Simulated users plus two public cohorts (364 real users).
- Simulated lead is uneven. On a 5,400-user simulated board the mini is slightly ahead of the full-size model; its lead over the Bayes reference comes from irregular cycles and is level or slightly behind on regular users.
- Missed logs hurt every method. Ask people if they forgot a period.
- Private by design. Runs locally, no accounts, nothing sent anywhere.