MyHeartCounts/openmhc-chronos2-fc

OpenMHC Forecasting — Chronos-2

Track 3 (forecasting) reference checkpoint for the MyHeartCounts / OpenMHC wearable-health benchmark (NeurIPS 2026).

This checkpoint is a Chronos-2 model. Chronos-2 is Amazon's universal time-series foundation model, supporting multivariate and covariate-informed probabilistic forecasting.

Fine-tuned from the pretrained amazon/chronos-2 base on the MHC training split (LoRA, rank 8, alpha 16; the adapter has been merged into the base so this is a standalone full model — no PEFT runtime dependency).

• Checkpoint format: HuggingFace model directory (checkpoint/: config.json + model.safetensors)
• Forecasting task: 24-hour-ahead, 19 sensor channels, hourly resolution.

Model & implementation

• Chronos (official implementation)
• Base model amazon/chronos-2
• Paper: Chronos-2: From Univariate to Universal Forecasting (Ansari et al., 2025)

Usage

import openmhc
from openmhc.forecasters import Chronos2Forecaster

# pip install "openmhc[chronos]"
fc = Chronos2Forecaster.from_release("hf://MyHeartCounts/openmhc-chronos2-fc@v1.0")
results = openmhc.evaluate_forecasting(fc, version="full")

The same bundle also loads in the evaluation harness via model.release_dir=hf://MyHeartCounts/openmhc-chronos2-fc@v1.0. See openmhc_manifest.json for provenance (training run, base model, fine-tuning details) and architecture metadata.

Citation

If you use this checkpoint, please cite the OpenMHC benchmark and the original Chronos-2 work (linked above).