ChiliGround - A verbatim RAG framework

A sentence classification model for extracting relevant spans from documents based on a question.

Model Details

• Base model: answerdotai/ModernBERT-base
• Hidden dimension: 768
• Number of labels: 2

Usage

from verbatim_rag.extractors import ModelSpanExtractor
from verbatim_rag.document import Document

# Initialize the extractor
extractor = ModelSpanExtractor(
    model_path="KRLabsOrg/chiliground-base-modernbert-v1",
    threshold=0.5
)

# Create documents
documents = [
    Document(
        content="""
        Climate change is a significant and lasting change in the statistical distribution of weather patterns.
        Global warming is the observed increase in the average temperature of the Earth's atmosphere and oceans.
        Greenhouse gases include water vapor, carbon dioxide, methane, nitrous oxide, and ozone.
        Human activities since the beginning of the Industrial Revolution have increased greenhouse gas levels.
        """,
        metadata={"source": "example_doc_1", "id": "climate_1"},
    ),
    Document(
        content="""
        Renewable energy comes from sources that are naturally replenished on a human timescale.
        Solar power is the conversion of energy from sunlight into electricity.
        Wind power is the use of wind to provide mechanical power or electricity.
        Hydropower is electricity generated from the energy of falling water.
        """,
        metadata={"source": "example_doc_2", "id": "energy_1"},
    ),
]


# Extract relevant spans
question = "What causes climate change?"
results = extractor.extract_spans(question, documents)

# Print the results
for doc_content, spans in results.items():
    for span in spans:
        print(span)

Training Data

This model was trained on a QA dataset to classify sentences as relevant or not relevant to a given question.

Limitations

• The model works at the sentence level and may miss relevant spans that cross sentence boundaries
• Performance depends on the quality and relevance of the training data
• The model is designed for English text only