federicosabbadini/patch-aliasing-models

Chronos-Bolt Patch-Stride Sweep

Retrained variants of Chronos-Bolt Tiny (~8.7M params) with different patch size (P) and patch stride (S) configurations, trained from scratch on the full official Chronos pre-training data.

Purpose

These models support the structural patch-aliasing study: investigating how the (P, S) patch geometry of Chronos-Bolt creates frequency-dependent artifacts in time-series forecasting. Each model is trained identically except for the patch geometry, so any downstream difference in aliasing probes is attributable to (P, S).

Models

Subfolder	P	S	Overlap	Status
p8-s4-seed42	8	4	0.500	done
p8-s8-seed42	8	8	0.000	done
p16-s4-seed42	16	4	0.750	done
p16-s8-seed42	16	8	0.500	done
p16-s12-seed42	16	12	0.250	done
p16-s15-seed42	16	15	0.062	done
p16-s16-seed42	16	16	0.000	done
p24-s8-seed42	24	8	0.667	done
p24-s12-seed42	24	12	0.500	done
p24-s15-seed42	24	15	0.375	done
p24-s16-seed42	24	16	0.333	done
p24-s20-seed42	24	20	0.167	done
p24-s24-seed42	24	24	0.000	done
p32-s8-seed42	32	8	0.750	done
p32-s12-seed42	32	12	0.625	done
p32-s15-seed42	32	15	0.531	done
p32-s16-seed42	32	16	0.500	done
p32-s20-seed42	32	20	0.375	done
p32-s24-seed42	32	24	0.250	done
p32-s28-seed42	32	28	0.125	done
p32-s32-seed42	32	32	0.000	done

Training setup

All models share the same training regime (only P and S vary):

• Architecture: Chronos-Bolt Tiny (T5-based, ~8.7M params), random initialization
• Data: Official Chronos pre-training corpus — TSMixup (10M series) + KernelSynth (1M series) at 9:1 ratio
• Steps: 100,000 (fixed budget across all runs)
• Optimizer: AdamW, lr=1e-3, linear decay, no warmup
• Batch size: 32
• Precision: fp32 + TF32 matmuls (Ampere+ GPUs)
• Context: 2048 tokens, prediction horizon: 64
• Quantiles: 9 (0.1 to 0.9)
• Seed: 42

Full provenance is recorded in each subfolder's run_config.json.

Usage

from chronos import BaseChronosPipeline

pipe = BaseChronosPipeline.from_pretrained(
    "federicosabbadini/chronos-bolt-patch-sweep",
    subfolder="p24-s16-seed42",
    device_map="cpu",
)

Files per model

• config.json — model architecture config
• model.safetensors — trained weights (~34 MB)
• run_config.json — full training provenance + result metrics
• loss_history.npy — per-step training loss
• loss_curve.png — training loss plot

License

The models are derived from the Chronos architecture (Apache-2.0) and trained on the official Chronos datasets. See the original Chronos repository for details.