[anishsaha12]

iAmTime is a 300M-parameter time-series foundation model for instruction-conditioned in-context learning, introduced in A Foundation Model for Instruction-Conditioned In-Context Time Series Tasks (code, weights).

Unlike conventional forecasting-only foundation models, iAmTime is trained on multiple time-series tasks using structured input-output demonstrations. Its hierarchical encoder and task-conditioned decoder enable adaptation to forecasting, classification, imputation, anomaly detection, and other tasks without parameter updates at inference time.

This PR adds:

• models/iamtime/model.py and models/iamtime/requirements.txt: an FEV-compatible wrapper for the public iAmTime forecasting checkpoint.
• benchmarks/fev_bench/results/iamtime.csv: evaluation results across all 100 fev-bench tasks.
• An iamtime entry in benchmarks/fev_bench/models.yaml.

Results

Skill scores are computed from the submitted CSV using FEV's leaderboard implementation and the Seasonal Naive baseline across the 100 tasks:

Model	SQL Skill (%)	MASE Skill (%)	WQL Skill (%)
iAmTime	46.8	35.4	51.6

The wrapper evaluates the pretrained checkpoint without task-specific fitting. The model returns probabilistic quantile forecasts and supports historical and future-known covariates.

Reproduction

uv run python models/evaluate.py \
  -m iamtime \
  -b benchmarks/fev_bench/tasks.yaml \
  -n iAmTime \
  -k '{"model_path":"anishsaha/iAmTime-base-forecast","device":"cuda","torch_dtype":"bfloat16","batch_size":64,"use_exogenous":true,"use_past_dynamic":true,"use_known_dynamic":true}'

[shchur]

According to Table 24 in the paper, some datasets present in fev-bench such as KDD 2022 were used for model pretraining. These should be disclosed in this field. Can you please update it to reflect which datasets among https://huggingface.co/datasets/autogluon/fev_datasets the model has seen during pretraining?

[anishsaha12]

Thanks @shchur for pointing this out! We had also noticed inconsistencies in the pretraining dataset list and had already corrected the tables to reflect the data actually used. The arXiv version hasn’t been updated yet, as the paper is currently under review, but the current version is available [on GitHub].

To clarify, none of the fev-bench datasets were used during pretraining, including KDD Cup 2022 and Favorita sales, which were incorrectly listed in the older version.

[shchur]

Thanks for the clarification, should be fine then.