Recordings and 4-way stroke category annotations of tabla playing of solo compositions and accompaniment to vocals (tabla recorded in isolation) released with the following conference paper: M. A. Rohit, A. Bhattacharjee, and P. Rao, “Four-way Classification of Tabla Strokes with Models Adapted from Automatic Drum Transcription”, in Proc. of the 22nd Int. Society for Music Information Retrieval Conf., Online, 2021. The dataset is split into test and train sets. The test set consists of 10 pieces of only the tabla accompaniment recorded in perfect isolation to prerecorded solo Hindustani vocal tracks. It contains 20 minutes of audio and nearly 4,500 strokes. These recordings, made on 3 unique tabla sets by 2 different artists, are diverse in terms of tuning, tala (metre), and tempo. The training set consists of solo compositions and common theka patterns recorded from 10 different tabla-sets. The total audio duration is about 1.25 hours and there are 26,600 strokes. The test set was annotated by first running an automatic onset detector to obtain stroke onsets, followed by manually assigning the four-way labels by listening to the audio and visually inspecting the spectrogram. The training set was annotated by automatically aligning the composition score (supplied by artists) with the audios, and replacing the bols with corresponding target stroke categories. Given the imperfect score-stroke matching, labels were manually verified to assign the same category to similar sounding bols. Onsets are separated into folders for each stroke category. Each '.onsets' file in these folders corresponds to an audio ('.wav') file of the same name and is a text file containing a list of time instants where an onset of a particular stroke category occurs.