jumafernandez/d2f-turn-embeddings-personachat

Turn embeddings for Personachat (Dialog2Flow encoder)

One 768-d float16 vector per utterance of Persona-Chat (Zhang et al., 2018), released within ParlAI, 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

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

Utterance text is not redistributed; recover it from the source corpus (https://github.com/facebookresearch/ParlAI, license mit) joining on the row order defined by the slim frame. The dataset was released by Facebook AI within the MIT-licensed ParlAI framework, with no separate dataset license; MIT is inherited from that release. Dialogues reconstructed from the truecased distribution (bavard/personachat_truecased) taking the longest history per conversation.

Load

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