---
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)