| # Atome LM v2 — SuperESP heads |
|
|
| Signed, on-device classifier "heads" for the [Atome LM](https://github.com/TilelliLab/atome-lm) |
| ternary (1.58-bit) engine. Each head runs **on the ESP32 itself**, bit-exact to the Python |
| reference, instead of generating text. |
|
|
| - **Heads:** agri, voice, motion, sound-scene, anomaly, air, os-telemetry, power/NILM, |
| occupancy, wearable, water, forecast (+ regression). |
| - **Files:** `*.atomecl` (signed head blob), `*.att.json` (Ed25519 attestation), |
| `*.tok.json` (tokenizer), `suite_pubkey.txt` (public verification key), |
| `RESULTS.json` / `PARITY.json` / `ATTESTATION.json` (held-out metrics + parity). |
| - **Verified:** 12/12 ran on a physical ESP32-WROOM-32. Several heads ship on clearly-labelled |
| physics-grounded synthetic data — retrain on your own data via the kit's `train --csv` path. |
| - **License:** Apache-2.0. Code + installer: https://github.com/TilelliLab/atome-lm (v2.0). |
|
|
