otacilio-psf's picture
Update README.md
935bff5 verified
---
dataset_info:
features:
- name: id
dtype: int64
- name: title
dtype: string
- name: ingredients
dtype: string
- name: directions
dtype: string
- name: link
dtype: string
- name: source
dtype: string
- name: NER
sequence: string
- name: metadata
struct:
- name: NER
sequence: string
- name: title
dtype: string
- name: document
dtype: string
- name: all-MiniLM-L6-v2
sequence: float32
- name: bm42-all-minilm-l6-v2-attentions
struct:
- name: indices
sequence: int64
- name: values
sequence: float64
splits:
- name: train
num_bytes: 1176543723
num_examples: 350000
download_size: 1101274243
dataset_size: 1176543723
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---
# Recipe Short - Dense and Sparse Embeddings Dataset
This dataset is based on the [rk404/recipe_short](https://huggingface.co/datasets/rk404/recipe_short) dataset, which itself is derived from the [RecipeNLG](https://recipenlg.cs.put.poznan.pl/) dataset. RecipeNLG is a large-scale, high-quality dataset designed for natural language generation tasks in the culinary domain. This dataset includes dense and sparse embeddings for each recipe, generated using the following models:
1. **Dense Embeddings**: Created using the `sentence-transformers/all-MiniLM-L6-v2` model with `fastembed` library.
2. **Sparse Embeddings**: Generated using the `Qdrant/bm25-all-minilm-l6-v2-attentions` model with `fastembed` library.
The embeddings were computed using GPU resources on Kaggle for efficient processing. This dataset is intended for tasks related to text similarity, search, and semantic information retrieval within recipe-related content.
### Sparse Embedding Model Reference
Sparse vector embedding model focuses on capturing the most important tokens from the text. It provides attention-based scores to highlight key terms, which can be beneficial for keyword-based search and sparse retrieval tasks.
You can find more about sparse embedding [here](https://qdrant.tech/articles/bm42/#:~:text=Despite%20all%20of%20its%20advantages,%20BM42) and [here](https://github.com/qdrant/bm42_eval/)
### Generation Code
[recipe-short-embeddings-gpu.ipynb](https://huggingface.co/datasets/otacilio-psf/recipe_short_dense_and_sparse_embeddings/blob/main/recipe-short-embeddings-gpu.ipynb)