vidore/baselines-v2

Contextualized Embeddings Benchmark

This repository contains evaluation code for the "Contextualized Embeddings" project.

Installation

pip install -e .
pip install git+https://github.com/jina-ai/late-chunking --no-deps  # for late-chunking with jina

Usage

Refer to scripts/evaluation.py for an example of how to use the code.

from datasets import load_dataset
from cde_benchmark.embedders.sentence_transformer_embedder import SentenceTransformerEmbedder
from cde_benchmark.embedders.naive_contextual_embedder import NaiveContextualEmbedder
from cde_benchmark.formatters.data_formatter import DataFormatter

# Datasets should be correctly formatted
formatter = DataFormatter("illuin-cde/chunked-mldr", split="test")

# Non-nested example
embedder = SentenceTransformerEmbedder("intfloat/e5-base-v2")
metrics = embedder.compute_metrics_e2e(formatter)
print(metrics)

# Nested example (for conxtualized embeddings models)
embedder = NaiveContextualEmbedder("intfloat/e5-base-v2")
metrics = embedder.compute_metrics_e2e(formatter)
print(metrics)