FLiP-data

Preprocessed data for the FLiP project — Factorized Linear Projection for Interpreting Multimodal Multilingual Sentence Embeddings.

FLiP trains a factorized log-linear model to recover lexical content (keywords) from pretrained sentence embeddings via a single linear projection, with no fine-tuning of the encoder.

Contents

SONAR embeddings and transcripts for Mozilla Common Voice v15 English (train / dev / test):

File	Description
*_speech_embs.npy	SONAR speech embeddings (float32, shape [N, 1024])
*_text_embs.npy	SONAR text embeddings (float32, shape [N, 1024])
*_sim_scores.npy	Cosine similarity between paired speech and text embeddings
*_transcript.txt	Reference transcripts (one utterance per line)
*_entities_gemini2.5_flash_lite.jsonl	Named entities extracted with Gemini 2.5 Flash Lite

Splits: train (1M utterances), dev (16k), test (~16k).

Source data

Embeddings were computed from Mozilla Common Voice v15 English using the SONAR encoder. Audio and transcripts from Common Voice are licensed under CC BY 4.0.

Trained checkpoints

HF repo	Training data	Embedding	Rank	Size
BUT-FIT/FLiP-en-sonar → mcv15/rank-512/	MCV v15 EN	SONAR	512	207 MB
BUT-FIT/FLiP-en-sonar → mcv15/rank-1024/	MCV v15 EN	SONAR	1024	414 MB

Usage

See the FLiP GitHub repo for full installation instructions and training/evaluation scripts.

Quick start after downloading:

import numpy as np

train_speech = np.load("cv_15/en/sonar_embeddings/train_speech_embs.npy")
train_text   = np.load("cv_15/en/sonar_embeddings/train_text_embs.npy")

Citation

@misc{kesiraju2026flip,
  title         = {{FLiP}: Towards understanding and interpreting multimodal multilingual sentence embeddings},
  author        = {Kesiraju, Santosh and Yusuf, Bolaji and Sedl{\'a}{\v{c}}ek, Simon and Plchot, Old{\v{r}}ich and Schwarz, Petr},
  year          = {2026},
  eprint        = {2604.18109},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2604.18109},
}