cryptolab-playground/Bloomberg-Financial-News-embedding-gemma-300m

Bloomberg Financial News Embeddings for Vector Database Benchmarking

Dataset Description

This dataset contains pre-computed embeddings of Bloomberg financial news articles, designed for evaluating vector database performance. The embeddings are generated using Google's EmbeddingGemma-300M model.

Purpose

Benchmark dataset for evaluating vector database performance on financial news domain, specifically designed for use with VectorDBBench.

Dataset Summary

• Total Training Samples: 368,816
• Test Queries: 1,000
• Ground Truth: Top-1000 nearest neighbors per query
• Embedding Dimension: 768
• Embedding Model: google/embeddinggemma-300m
• Source Data: danidanou/Bloomberg_Financial_News

Dataset Structure

Data Splits

Split	Samples	Description
train	368,816	Training embeddings (80% random sample from source)
test	1,000	Test query embeddings (from remaining 20%, non-overlapping)
neigbors.parquet	1,000	Top-1000 nearest neighbors for each test query

Data Fields

train & test

• id (int64): Unique identifier for each article
• emb (List[float64]): 768-dimensional L2-normalized embedding vector

neigbors.parquet

• id (int64): Query identifier (matches test)
• neighbors_id (List[int64]): List of 1000 nearest neighbor IDs from train set

Dataset Creation

Source Data

The dataset is derived from approximately 447K Bloomberg financial news articles:

• Train: 80% random sample (368,816 articles)
• Test: 1,000 articles randomly sampled from remaining 20% (non-overlapping with train)

Preprocessing

1. Text Preparation: Concatenated Headline + Article for each news item
2. Chunking: For texts exceeding 2048 tokens:
   • Split into chunks with ~100 token overlap
   • Embedded each chunk separately
   • Averaged chunk embeddings for final representation
3. Normalization: All embeddings are L2-normalized

Embedding Generation

• Model: google/embeddinggemma-300m
• Dimension: 768
• Max Token Length: 2048
• Normalization: L2-normalized

Ground Truth Generation

Ground truth nearest neighbors were computed using:

• Method: Flat search (brute-force)
• Metric: Cosine similarity
• K: Top-1000 neighbors per query

Usage

Loading the Dataset

from datasets import load_dataset
import pandas as pd

# Load train and test splits
dataset = load_dataset("redcourage/Bloomberg-Financial-News-embedding-gemma-300m")
train = dataset['train']
test = dataset['test']

# Load ground truth
neigbors = pd.read_parquet(
    "hf://datasets/redcourage/Bloomberg-Financial-News-embedding-gemma-300m/neigbors.parquet"
)

Evaluation Example

import numpy as np
from datasets import load_dataset
import pandas as pd

# Load data
dataset = load_dataset("redcourage/Bloomberg-Financial-News-embedding-gemma-300m")
train_data = dataset['train']
test_data = dataset['test']
neigbors = pd.read_parquet(
    "hf://datasets/redcourage/Bloomberg-Financial-News-embedding-gemma-300m/neigbors.parquet"
)

# Convert to numpy arrays
train_embeddings = np.array(train_data['emb'])
test_embeddings = np.array(test_data['emb'])

# Example: Compute recall@10
def compute_recall_at_k(retrieved_ids, neigbors_ids, k=10):
    """
    Compute Recall@K
    
    Args:
        retrieved_ids: List of retrieved neighbor IDs
        neigbors_ids: List of ground truth neighbor IDs
        k: Number of top results to consider
    """
    retrieved_k = set(retrieved_ids[:k])
    neigbors_k = set(neigbors_ids[:k])
    
    if len(neigbors_k) == 0:
        return 0.0
    
    return len(retrieved_k & neigbors_k) / len(neigbors_k)

# Use with your vector database
# ... insert your vector DB search code here ...

Use Cases

• Vector database performance benchmarking on financial domain
• Approximate nearest neighbor (ANN) algorithm evaluation
• Retrieval system testing for financial news

Limitations

• Domain-Specific: Optimized for financial news; may not generalize to other domains
• Language: English only
• Temporal Coverage: Limited to articles available in the source dataset (2006-2021)
• Chunking Strategy: Long documents are averaged, which may lose fine-grained information
• Ground Truth: Based on cosine similarity with embeddings, not human relevance judgments
• Financial Bias: May reflect biases present in Bloomberg's reporting and article selection

License

Apache 2.0

Citation

If you use this dataset, please cite:

@dataset{bloomberg_embeddings_gemma,
  author = {redcourage},
  title = {Bloomberg Financial News Embeddings for Vector Database Benchmarking},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/redcourage/Bloomberg-Financial-News-embedding-gemma-300m}
}

Source Dataset Citation

@dataset{bloomberg_financial_news,
  author = {danidanou},
  title = {Bloomberg Financial News},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/danidanou/Bloomberg_Financial_News}
}

Embedding Model Citation

@misc{embeddinggemma,
  title={Embedding Gemma},
  author={Google},
  year={2024},
  url={https://huggingface.co/google/embeddinggemma-300m}
}

Acknowledgments

• Original dataset: danidanou/Bloomberg_Financial_News
• Embedding model: google/embeddinggemma-300m
• Benchmark framework: VectorDBBench