Family
TimesFM 2.0 (500M params, LoRA fine-tune, PyTorch)
in-developmentApache 2.0tabularretail-facing
Goal Attainment Forecaster (TimesFM 2.0).
Saver Room — quantile forecast of goal attainment progress (p10/p50/p90 over 30d/90d) via TimesFM 2.0 LoRA fine-tune.
Foundation · TimesFM 2.0 (500M params, LoRA fine-tune, PyTorch)
What it does
The honest description.
Saver Room — quantile forecast of goal attainment progress (p10/p50/p90 over 30d/90d) via TimesFM 2.0 LoRA fine-tune.
Architecture
How it works.
Per-user signed_amount window (PR-2a stage 8 timesfm_windows.parquet output) — 256-day context length, daily resample.sum.fillna(0). TimesFM 2.0 backbone weights + LoRA adapter (PR-3 artifact). Quantile forecast (p10/p50/p90) over 30d/90d horizons + isotonic recalibration applied + monotone-repair clamped (TabularRisk F1 invariant). ForecastResponse Pydantic shape (PR-5).
Training data & drift
What it learned from, and how we watch it.
Source
Inputs: Per-user signed_amount window (PR-2a stage 8 timesfm_windows.parquet output) — 256-day context length, daily resample.sum.fillna(0). TimesFM 2.0 backbone weights + LoRA adapter (PR-3 artifact).
Drift baseline · current quarter
Drift baseline not yet documented for this model.
What it's NOT
- Operator-gated to first AWS GPU rollout — TimesFmTritonAdapter raises ConfigurationError until TRITON_FORECASTER_URL is set + Triton forecast-pool deployed.
Version history · immutable once published
Every release, traceable.
Data Mukodi holds about you is subject to DPDP §12 erasure (use /consent in the app), and any concern about this model can be raised via [email protected] under DPDP §13 grievance, or via /security.