TIMBRE: Teaching Time Series Forecasters to Read, Remember, and Reconcile

Temporal Integration of Memory-Based Responses and Evidence

Clean experiment code · Loading instructions

This repository contains the final trained weights from the completed 2026-09-24 development experiment:

Directory	Contents
full/	Final FULL evidence adapter, model identity and fixed training configuration
ordinary_esm/	Final matched ordinary-fusion baseline adapter and configuration
memory/	17 final FIT-prefix memory snapshots and minimal loading/cutoff metadata

There are 19 weight files, approximately 18 MB in total. No optimizer state, intermediate checkpoint, training log, prediction file or dataset is included. The frozen base model is downloaded separately from amazon/chronos-2 at revision 29ec3766d36d6f73f0696f85560a422f50e8498c.

Training scope

FULL and ordinary fusion used seed 17, one epoch, 13,793 FIT windows and 432 updates. They were evaluated on 2,239 TUNE windows across four data families and 13 tasks. The adapters retain the final update; no checkpoint was selected by TUNE performance. Each memory group uses the earliest 20 percent of FIT, 256 updates, and a group-specific availability cutoff.

These weights are a development release. Retrained ablations, multiple seeds and independent final-test evaluation remain incomplete. The current results do not establish overall superiority to the frozen numeric base. Recorded availability clocks and the semantic validity of auxiliary scores remain uncertified. Missing scores and lead-specific memory support must remain masked.

Loading

Clone the linked code repository and install its pinned requirements with Python 3.11. Follow its README to load full/adapter.pt or ordinary_esm/adapter.pt with FixedEvidenceForecaster.load_adapter and the matching frozen backbone. These are custom TIMBRE adapters, not AutoModel or PEFT checkpoints. Loading uses torch.load(weights_only=True) and validates the backbone, configuration, postprocessor and tensor membership.

Before loading, set backbone.config._name_or_path = "amazon/chronos-2". The original adapter configuration digest included a local absolute path. The published digest uses this stable model identity instead; all parameter tensors are exactly unchanged. WEIGHTS_MANIFEST.json records both original and published checkpoint hashes. The memory files are byte-identical to the final trained snapshots.

The adapter expects a validated EvidenceBatch and its corresponding numeric cache. The base model, fixed 1024-dimensional text features, evidence records and historical donor data are not embedded in the adapter. Full input contracts and the distinction between accepted bundles and the historical development loader are documented in the code repository.

For memory loading, use memory/SNAPSHOTS.json with timeline_v2.fixed_memory.learn.load_model. Respect each group's available_at timestamp and point-wise support mask.

Verification

Both adapters and all 17 memory snapshots were reloaded using the clean source release. Adapter tensors were compared with the original final training checkpoints using exact equality; memory files were checked by SHA256. The scoped source suite ran 253 tests, with 252 passing and one data-dependent integration test skipped. The original FULL run's independent result verifier also passed without rerunning training.

WEIGHTS_MANIFEST.json binds the code release and every payload file. SHA256SUMS provides file-level checksums.

Code revision: ecdeee9.