FLiP-data
收藏资源简介:
FLiP-data是为FLiP项目预处理的数据集,旨在通过因子化线性投影解释多模态多语言句子嵌入。数据集包含Mozilla Common Voice v15英语的SONAR语音和文本嵌入(形状为[N, 1024]的float32数组)、转录文本以及语音和文本嵌入之间的余弦相似度分数。此外,还包含通过Gemini 2.5 Flash Lite提取的命名实体。数据集分为训练集(约100万条语句)、开发集(约1.6万条)和测试集(约1.6万条)。数据来源于Mozilla Common Voice v15,使用SONAR编码器计算嵌入,音频和转录文本遵循CC BY 4.0许可。适用于句子相似性和特征提取等任务。
FLiP-data is a preprocessed dataset for the FLiP project, aiming to explain multimodal multilingual sentence embeddings through factorized linear projections. The dataset includes SONAR speech and text embeddings (float32 arrays of shape [N, 1024]), transcriptions, and cosine similarity scores between speech and text embeddings from Mozilla Common Voice v15 English. Additionally, it contains named entities extracted via Gemini 2.5 Flash Lite. The dataset is divided into a training set (approximately 1 million utterances), a development set (about 16,000), and a test set (about 16,000). The data originates from Mozilla Common Voice v15, with embeddings computed using the SONAR encoder, and the audio and transcriptions follow the CC BY 4.0 license. It is suitable for tasks such as sentence similarity and feature extraction.
FLiP-data 数据集详情
数据集概述
- 名称:FLiP-data
- 项目:FLiP(Factorized Linear Projection for Interpreting Multimodal Multilingual Sentence Embeddings)
- 语言:英语(单语)
- 许可证:CC BY 4.0
- 任务类别:句子相似度、特征提取
- 标签:SONAR、语音嵌入、文本嵌入、Common Voice、可解释性
数据集内容
该数据集包含 Mozilla Common Voice v15 英语 数据集的 SONAR 嵌入和转录文本,分为训练集、开发集和测试集。
| 文件名称 | 描述 |
|---|---|
*_speech_embs.npy |
SONAR 语音嵌入(float32,形状 [N, 1024]) |
*_text_embs.npy |
SONAR 文本嵌入(float32,形状 [N, 1024]) |
*_sim_scores.npy |
配对语音与文本嵌入之间的余弦相似度 |
*_transcript.txt |
参考转录文本(每行一条语句) |
*_entities_gemini2.5_flash_lite.jsonl |
使用 Gemini 2.5 Flash Lite 提取的命名实体 |
数据集划分
- 训练集:约 100 万条语句
- 开发集:约 1.6 万条语句
- 测试集:约 1.6 万条语句
数据来源
- 嵌入由 SONAR 编码器从 Mozilla Common Voice v15 英语数据计算得出。
- Common Voice 的音频和转录文本遵循 CC BY 4.0 许可。
训练检查点
| Hugging Face 仓库 | 训练数据 | 嵌入 | 秩 | 大小 |
|---|---|---|---|---|
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 |
使用示例
python 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")
参考论文
- 论文标题:FLiP: Towards understanding and interpreting multimodal multilingual sentence embeddings
- arXiv:2604.18109




