autokai/chronos2-fullft-dlam-g100

Chronos-2, fully fine-tuned (TU Darmstadt DLAM SoSe 2026 bonus project, group 100)

A full fine-tune (not LoRA) of amazon/chronos-2 on the course's 96-series operational-load panel, fitted on all 4320 labelled hours (_hidx < 4320, i.e. strictly the observed region).

It exists because the graded submission archive is capped at 200 MB and a full fine-tune is 455.8 MiB — the whole model, since full FT rewrites every weight, against LoRA's 4.6 MiB adapter. Hosting it here and fetching it at inference is the route the course staff sanctioned.

• Role: a blend MEMBER at weight 0.4245, never a covariate.
• Blend: 0.4245 * this + 0.4482 * TFT-cascade-bag5 + 0.1273 * LightGBM -> pooled far-regime WAPE 0.12722 over three gapped CV windows, against 0.13163 for the two-member predecessor.
• Training: finetune_mode=full, lr 1e-5, 1000 steps, batch 96, context 1024, prediction_length 672, seed 42.
• Reproduce: python -m src.models.chronos2_finetune --config configs/chronos2_fullft.yaml --cut-idx 4320 --no-eval

Fitted on course-provided data for a course submission; not a general-purpose forecaster.