acceptodds
Withdrawn from ICLR 2027

TimeNet: An Extensible Unified Data Infrastructure for Next-Generation Temporal Foundation Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Temporal Foundation Models (TFMs) aim to generalize across domains, datasets, and tasks. Yet, their development remains constrained by fragmented, task-specific data formats, annotations, and processing pipelines. We introduce TimeNet, an open-source data standard and scalable infrastructure that decouples temporal data from task definitions and represents signals, metadata, annotations, and supervision in a shared, extensible data model. TimeNet supports multimodal signals with regular, irregular, or ordinal time axes and expresses different task families (including classification, forecasting, temporal localization, question answering, generation, and editing) as reusable views over the same recordings. This shared representation enables heterogeneous time-series datasets to be combined for large-scale model training across domains, modalities, and tasks. We demonstrate TimeNet by transcoding datasets with 1.5M task instances spanning diverse domains, modalities, temporal scales, and forms of supervision, while retaining practical I/O performance relative to native formats. TimeNet enables a TFM training pipeline to support joint training on a configurable number of heterogeneous datasets through configuration changes alone. We show this capability by training a Time Series Language Model (TSLM) across multiple datasets and tasks, obtaining a 14% F1 score improvement compared with models trained on individual datasets. These results show that TimeNet provides the data and systems foundation needed to move beyond task- and dataset-specific TFMs toward models that can learn jointly across heterogeneous domains, modalities, temporal scales, and forms of supervision from a common data model.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Time-series foundation models, multimodal data infrastructure, unified data standard, multimodal time series, multitask learning, dataset interoperability, benchmarking

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Ek8bOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces TimeNet, a unified data standard and infrastructure for foundation models, aimed at addressing issues such as the task-specific nature of data resources, fragmented data formats, and the redundant implementation of data loaders.

Critical strengths

1、The problem addressed is both real and significant. In the time-series domain, common challenges include task-specific data packaging, the redundant repackaging of the same records, and the repeated implementation of data loaders. The argument for decoupling data from tasks is clearly articulated.

2、The data model design is quite comprehensive. Several design elements offer tangible engineering value, including the recursive hierarchy of Records, Sources, and Signals; the support for three distinct types of axes (regular, irregular, and ordinal); a unified approach to defining Spans; the separation of annotation content from its specific occurrences; and the use of task-based references to annotation occurrences, which avoids unnecessary data duplication.

Critical weaknesses

1、The substantive differences between the method proposed in this paper and existing data standards or formats require further clarification. While the paper references works such as Löning et al. (2019), Alexandrov et al. (2020), Herzen et al. (2022), Prabowo et al. (2024), and AtriumDB, the main text fails to sufficiently articulate the unique, indispensable value that "TimeNet" offers over these established standards. For instance, what is the core contribution of TimeNet when compared to general-purpose data standards such as the Arrow/Parquet ecosystem, WebDataset, TensorFlow Datasets, HuggingFace Datasets, or MLCommons Croissant? Furthermore, beyond features like support for multi-modality, diverse tasks, text-based supervision, and a versioned registry, how does TimeNet’s data model differ fundamentally from that of AtriumDB? While using DuckDB combined with Parquet is a logical engineering choice, one must ask: does this constitute a contribution at the level of a "data standard," or is it merely a "specific implementation"?

2、The value of task formalization has yet to be fully validated through experiments. While the concepts of "task signatures" and the classification of tasks into five distinct families are theoretically appealing, the current experiments are limited to QA and reasoning tasks. They fail to demonstrate, for instance, whether a unified collation process—applied to a dataset containing a mix of classification, forecasting, localization, and editing tasks—can truly function without the need for dataset-specific code. Furthermore, it remains unproven whether the proposed partial ordering for task specialization is useful in practical task organization, or whether joint training across different task families yields results that are superior (or at least not inferior) to single-task training. Therefore, it is essential to include an experiment—such as joint training or unified data loading across diverse task families (e.g., classification, forecasting, and localization)—to substantiate that this task framework is more than just a theoretical construct.

3、The empirical evidence for "pooled training" is relatively weak, and the strength of the conclusions should be tempered. For instance, the study covers only four QA datasets and two model families; furthermore, SLIP shows a notable performance decline on the ECG-QA dataset (F1: 33.93 → 30.00; Acc: 40.01 → 31.32), suggesting that pooled training does not consistently yield benefits. Additionally, the study fails to report variance, confidence intervals, or the number of random seeds used, leaving it unclear whether the observed improvements in F1 scores are statistically significant. While it is stated that the single-dataset models and the pooled models employ the "same architecture and training configuration," it remains ambiguous whether the total number of training steps, the total number of tokens, and the data sampling ratios are also identical. If the pooled models were exposed to a larger volume of data, the observed performance gains might stem from the increased data quantity rather than from the development of a "unified representation."

4、Details regarding reproducibility and registry governance are insufficient. For instance, what is the basis for the statistic of 1.51 million task instances? Do figures such as 899k for QA and 484k for classification involve double-counting across different datasets? Furthermore, how are aspects like registry versioning, licensing, access control, data deduplication, and train/val/test splits managed to ensure no data leakage into downstream evaluations? The "local reconstruction" of restricted data relies on connectors; however, how are the maintenance, version compatibility, and evolution of source data handled for these connectors? While anonymized code is provided as supplementary material, the main text fails to address test coverage, schema migration, or backward compatibility strategies.

Questions for authors

see weaknesses.

Reviewer_h8NqOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper titled “TimeNet: An Extensible Unified Data Infrastructure for Next-Generation Temporal Foundation Models” proposes a new model Timenet for foundational time series (TFM) which supports multimodal signals with regular, irregular, or ordinal time axes and expresses different task families (including classification, forecasting, temporal localization, question answering, generation, and editing) as reusable views over the same recordings. It alleviates the limitations of current TFMs in generalizing across domains, modalities, and tasks. The experiments show the value of TimeNet in transcoding datasets with 1.5M task instances and showing improvements in F1 score improvement compared with models trained on individual datasets.

Critical strengths

The pipeline for this approach decouples the data and task which seems to be beneficial. Storing continuous waveforms once and then layering multiple annotations, questions, forecasts,etc as relational views drastically reduces the storage footprint and preprocessing fragmentation. Combining DuckDB for the relational control plane with sharded Parquet files for is a good engineering design choice. The method also shows multi-dataset handling ability, with just doing some configuration changes.

Critical weaknesses

The proposed approach is a giant software engineering project. The deep learning models and training techniques employed are existing architectures adapted to ingest data from the unified interface. Consequently, the paper reads primarily as a data systems and standardization pipeline, making it far more suitable for a Datasets and Benchmarks track or a systems venue rather than a core machine learning research track. The experimental demonstration does not substantiate the broad applicability claims of the paper. The actual multi-dataset training evaluation is restricted almost entirely to text-based Question Answering and reasoning tasks using language-based models. As mentioned in the limitations, the data specification is currently static and offline. This part needs to be completed, before the paper is ready for publication.

Questions for authors

na

Reviewer_GKNTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces TimeNet, an open-source data standard and infrastructure that decouples temporal data from task definitions for Temporal Foundation Models (TFMs). Its core is the TimeF data model: datasets consist of Records owning recursive Source hierarchies containing Signals (with regular/irregular/ordinal axes); Spans address temporal points/intervals; Annotations are normalized, reusable typed observations attachable to any object; Tasks reference ordered tuples of records with targets stored inline or via annotation references. A compositional task taxonomy derives task families from a signature ϕ = (κ, Δ, α): output kind, target position, and arity. This yields five semantic families (attribute inference, temporal localization, future-directed inference, relational inference, temporal construction/transformation), realized as eight built-in task types. Storage separates a DuckDB control plane from sharded Parquet value planes. A registry hosts converted datasets as versioned org/name@version entries (currently 1.51M task instances: 899k QA, 484k classification, 131k generation, etc.). The experiments are of two kinds: (i) pooled multi-dataset training of OpenTSLM (Llama-3.2-1B Flamingo) and SLIP (400M) across four QA datasets, where OpenTSLM's average F1 improves from 29.16 to 33.28; and (ii) I/O benchmarks against native formats on six datasets, where performance is broadly comparable, sometimes better and sometimes worse.

Critical strengths
  1. A well-designed, actually-deployed standard with a working ecosystem. TimeF's separation of data from tasks, normalized annotation storage, versioned registry with 1.5M task instances, and open-source release give the contribution real adoption potential. Infrastructure papers live or die by whether the artifact exists and works; this one appears to.
  2. The compositional task taxonomy. The ϕ = (κ, Δ, α) signature cleanly factorizes "what/where/how many" of temporal tasks and yields a partition of the signature space. This is the kind of conceptual scaffolding a fragmented field needs, and it is directly operationalized in the software.
Critical weaknesses
  1. The empirical evidence for the central training claim is thin and partly negative.** Joint training is evaluated on one task family (QA) with two model families on four datasets; one of the two models loses substantially on one dataset, and the abstract's 14% figure reflects only the favorable model. For a paper whose pitch is "the data foundation for next-generation TFMs," the absence of any forecasting/classification/generation training experiment, and of multi-seed statistics, is a material gap.
  2. Insufficient differentiation from existing standards and infrastructure. GluonTS/sktime/darts interfaces, LOTSA/UTSD-style harmonized corpora, and medical event data standards (e.g., MEDS, cited) each cover part of this space. The paper would be much stronger with an explicit feature-level comparison (task model, annotation model, versioning, irregular axes, text supervision) and an analysis of what cannot be expressed in prior schemas.
Questions for authors
  1. Can you provide multi-seed results (or at least training curves) for the pooled-vs-single comparison, including the SLIP/ECG-QA regression? Is the regression reproducible, and is it really capacity-driven?
  2. What is the per-dataset engineering cost of writing a connector, and how much of the 1.5M-instance registry was built by the authors vs. community? This bears on the M+N scalability claim in practice.
  3. For the I/O cases where TimeF is markedly slower (Sleep-EDF throughput, request counts), what is the bottleneck? In particular, is random-access (window-level) performance, which is the regime that matters for training, affected the same way as full sequential reads?
  4. An explicit comparison table against GluonTS/sktime/MEDS-style schemas would help position the contribution.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.