Tabby: An Open Pretraining Recipe for Time Series Foundation Models
Shifeng Xie, Bahaeddine Abdessalem, Zehao Xiao, Youssef Attia El Hili, Ambroise Odonnat, Zhiwei Dong, Lei Zan, Themis Palpanas, et al.
cs.LG
Quick verdict
Tabby: open-recipe 145M time-series foundation model; strong on in-distribution GIFT-Eval, mid-table on OOD TIME; unusually transparent, well-ablated recipe
Summary
Tabby is a 145.843M-parameter time series foundation model handling contexts up to 8,192 and serving forecasting, classification, and anomaly detection from one backbone. An encoder-only patch Transformer is trained with masked quantile reconstruction on real data (GIFT-Eval-Pretrain+, BLAST) plus synthetic streams: KernelSynth and CauKerV2, a new generator composing temporal dynamics via sampled structural causal models online during training. Training contributions are a progressive convergence schedule yielding reusable intermediate checkpoints, and deep quantile supervision with trajectory regularization for intermediate layers. A prompt-tuning module (160 learnable tokens, frozen backbone) improves in-distribution forecasting. On GIFT-Eval, Tabby ranks third of eleven recent systems by MASE; on the out-of-distribution TIME benchmark sixth (MASE) and seventh (CRPS). UCR classification averages 0.7999 accuracy, edging Catch22+ by 0.003 without statistical testing; TSB-AD-U anomaly detection reaches 0.4282 VUS-PR, above zero-shot peers but below specialized methods like TSPulse. Most ablations use a 40M reference model, and some evaluation sections were truncated in the available text.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single pretrained 145.843M-parameter backbone (context ≤8,192) serves zero-shot forecasting, classification (UCR), and anomaly detection (TSB-AD-U).
- The pretraining corpus — GIFT-Eval-Pretrain+ and BLAST real data plus KernelSynth and the new online SCM-based generator CauKerV2 — is released with a complete construction recipe.
- The progressive convergence schedule yields locally converged, reusable intermediate checkpoints (55K/110K/165K) usable for evaluation and continuation.
- Deep quantile supervision improves the model: layer-bypass ablations (Appendix E) and norm-matched gradient interventions (Appendix G, ΔBM CIs strictly positive) support a directional benefit at intermediate depth.
- A prompt-tuning module (160 learnable tokens, frozen backbone) further improves in-distribution GIFT-Eval forecasting (0.698 vs 0.719 MASE for Tabby-Pretrain).
Strengths
Reproducibility is treated as a first-class contribution
The paper releases the full pipeline (corpus construction, generators, distributed training, checkpoint continuation, evaluation) and documents consequential choices usually omitted: sub-dataset-balanced sampling (Section 5.1), masking (Section 6), optimization (Table 16 even reports 40,291,456 parameters and 153.700 MiB checkpoint size for Tabby-Small), plus detailed corpus statistics in Appendix B.
Unusually honest evaluation framing
The authors explicitly label GIFT-Eval as in-distribution (corpus derived from GIFT-Eval-Pretrain+ and prompt-tuned on its training split), report Tabby-Pretrain separately from the prompt-tuned model in Figure 2, and concede TSPulse (0.48) beats Tabby (0.4282) on TSB-AD-U in Section 8.3.
CauKerV2 is a sensible synthetic-data contribution
Appendix-level analysis is well above average rigor
Clean anomaly-detection protocol
Weaknesses
Ranking language outranks the numbers
Key margins are tiny and untested; internal inconsistencies exist
The 'open recipe' is incomplete for the headline model
All component ablations run at proxy scale, single-seed
Minor coherence lapses
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Tabby (arXiv 2609.13956) is positioned less as a novel architecture than as an *open recipe*: an encoder-only patch Transformer (145.843M parameters, context ≤8,192) whose contributions concentrate on data (GIFT-Eval-Pretrain+ + BLAST real corpora; KernelSynth + the new SCM-based online generator CauKerV2) and training procedure (Progressive Convergence schedule, Deep Quantile Supervision with trajectory regularization, and a frozen-backbone prompt-tuning module of 160 learnable tokens). The headline claims: one backbone serving forecasting, classification, and anomaly detection; 'top tier zero-shot forecasting performance on GIFT-Eval and TIME'; and end-to-end open-source reproducibility.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Tabby and what tasks does it support?
Tabby is an open-source time series foundation model from Huawei Noah's Ark Lab (145.843M parameters, context up to 8,192 observations). One pretrained backbone supports three tasks: probabilistic quantile forecasting, time series classification (via frozen-encoder representations), and zero-shot anomaly detection (via Mahalanobis distance on patch embeddings). Code, training pipeline and model are released at huawei-noah/trustworthyAI.
What is CauKerV2?
CauKerV2 is Tabby's synthetic pretraining data generator. It extends the earlier CauKer generator from classification to forecasting: it draws primitive series from ten families of temporal dynamics and combines them through randomly sampled structural causal models, generating samples online during training so the model never overfits a fixed synthetic set.
How well does Tabby perform compared with other time series foundation models?
On the in-distribution GIFT-Eval benchmark it ranks 3rd of 11 recent systems by MASE (0.698). On the out-of-distribution TIME benchmark it is 6th by MASE (0.674) and 7th by CRPS (0.565), behind leaders like Sundial (0.642 MASE). For classification on UCR it reports 0.7999 mean accuracy — the best listed, but only 0.003 above Catch22+. On TSB-AD-U anomaly detection it reaches 0.4282 VUS-PR, above zero-shot peers (MOMENT 0.38, TimesFM 0.30) but below task-specialized methods such as TSPulse (0.48).
Is Tabby's code and model actually available?
Yes — the paper states that the training pipeline, the corpus construction recipe and generators, and the model itself are released as open source under the huawei-noah/trustworthyAI repository, explicitly for end-to-end reproduction.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The extracted text is truncated mid-paper: Section 6 (masking/normalization), parts of Section 9's ablation results, and several appendix analyses were omitted, so some claims could not be checked against the full text.
- Figures 1–3 (including the Figure 2 leaderboard values used for ranking assessments) were recoverable only from embedded axis labels and value lists, not from rendered figures.
- The metadata's author list (8 names) omits Jianfeng Zhang, Lujia Pan, Keli Zhang and Malik Tiomoko, who appear as authors in the paper text itself.
- Several numeric inconsistencies (Table 3's 0.430 vs 0.4282; Figure 1's 30/30/30/10 vs Table 16's 0.25/0.35/0.30/0.10) could not be re-verified against a complete PDF.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.