jumafernandez/d2f-turn-embeddings-taskmaster

Turn embeddings for Taskmaster (Dialog2Flow encoder)

One 768-d float16 vector per utterance of Taskmaster-1/2/3 (Google Research), computed with the frozen encoder sergioburdisso/dialog2flow-joint-bert-base (SentenceTransformer recipe, convert_to_numpy, no normalization). The encoder truncates inputs at 64 tokens. If you use these embeddings, please cite the Dialog2Flow paper (Burdisso et al., EMNLP 2024) and the source corpus.

Files

• taskmaster_e_t.f16.npy — numpy array (n_turns, 768), float16; row i is turn i.
• taskmaster_dialogs.slim.pkl — pandas DataFrame aligned row-by-row with the array: columns dataset, split, dialogue_id, turn_id, speaker (no utterance text).
• taskmaster_e_t.f16.npy.meta.json — encoding metadata.

Utterance text is not redistributed; recover it from the source corpus (https://github.com/google-research-datasets/Taskmaster, license cc-by-4.0) joining on the row order defined by the slim frame. TM-1, TM-2 and TM-3 combined; speaker roles mapped by name (USER/ASSISTANT).

Load

import numpy as np, pandas as pd
emb = np.load("taskmaster_e_t.f16.npy", mmap_mode="r")
meta = pd.read_pickle("taskmaster_dialogs.slim.pkl")
assert len(meta) == len(emb)