TimeBraid-SFT

Instruction-tuning data accompanying TimeBraid: Unifying Time Series and Language for Understanding and Forecasting. Stage 2 combines time-series forecasting, understanding, and description with general-language examples and selected alignment examples retained during training.

The full Stage-2 training inventory contained 4,881,583 examples (about 4.88M), distributed as follows:

Category	Source or subcategory	Examples
Forecasting		2,640,566
	CGTSF	53,926
	FinMultiTime S&P 500	144,387
	MoTime (News/Wiki)	32,589
	CAF-7M	2,314,216
	Time-MMD	20,768
	Time-MQA Forecasting	74,680
Time-series QA and reasoning		1,065,021
	CaTS-Bench	15,995
	ChatTS	50,377
	HiTSR	157,147
	OpenSQA	138,754
	OpenTSLM ECG	159,313
	OpenTSLM HAR	68,542
	OpenTSLM Sleep	7,434
	OpenTSLM TSQA	38,400
	Time-MQA Reasoning	110,212
	TSAQA Reasoning	125,957
	EngineMT-QA	100,296
	RATs40K	32,994
	Time-MQA Imputation	38,607
	TSAQA Anomaly Detection	20,993
Captioning and description		161,883
	OpenTSLM M4 Captioning	80,000
	SensorCaps	35,960
	Time-MMD Open-Ended	1,925
	TRACE Captioning	43,998
Unimodal corpora		400,000
	GIFT-Eval Pretrain	200,000
	Dolci-Instruct (No-Tools)	200,000
Alignment retention		614,113
	Understanding retention	116,294
	Forecasting retention	497,819
Total		4,881,583

These are training-example counts, not counts of unique time series. The alignment-retention examples are included in the training total; they will not be packaged again in this dataset alongside the companion TimeBraid-Alignment collection.

Coming soon

Dataset files are not available here while we review the licenses and redistribution permissions of the underlying sources. The distribution above describes the full training inventory, not a downloadable package; the final shared subset may differ. We will update this page when we know what can be shared.

Stay tuned.