truthscore 0.3.0

A fast, modular reimplementation of RAGAS's FactualCorrectness metric, supporting both open-weight and dedicated LLMs.

TruthScore

``truthscore` is a fast, modular reimplementation of RAGAS's FactualCorrectness metric, supporting both open-weight and hosted LLMs. It evaluates factual consistency between a user response and a reference passage by breaking down answers into claims and verifying them using Natural Language Inference (NLI).

It is a component of the trutheval framework and is intended for scalable, cost-efficient factuality evaluation.

---

🔍 What it does

1. Claim Decomposition: The LLM-generated response is split into atomic factual claims using a lightweight LLM.
2. Entailment Scoring: Each claim is passed to an NLI model with the reference passage as context.
3. Final Score: The score reflects how many claims are entailed by the context, in the range [0.0, 1.0].

For more details, see FactualCorrectness.

---

✨ Key Features

• 🔁 RAGAS-compatible: Faithfully reimplements the FactualCorrectness metric logic from RAGAS
• ✅ Open-weight LLM support: Works with open-weight models (e.g., Gemma, LLaMA, Mistral via Ollama)
• 🧠 Plug-and-play: Swap in custom NLI models
• ⚙️ GPU-accelerated: Recommended for claim decomposition + NLI
• 🧪 Evaluated: Competitive benchmark results (see trutheval)

---

📦 Installation

For full open-weight support (LLM hosted with Ollama + CrossEncoders NLI):

pip install truthscore[open]

Otherwise, use the lightweight dependency pick the dependencies that best work for you with:

pip install truthscore

Regarding ollama installation, please check Ollama.

🚀 Quick Start

💡 Open-weight (fully local)

from langchain_community.llms import OllamaLLM
from ragas import SingleTurnSample
from ragas.llms import LangchainLLMWrapper

from truthscore import OpenFactualCorrectness

test_data = {
    "user_input": "What happened in Q3 2024?",
    "reference": "The company saw an 8% rise in Q3 2024, driven by strong marketing and product efforts.",
    "response": "The company experienced an 8% increase in Q3 2024 due to effective marketing strategies and product efforts."
}
sample = SingleTurnSample(**test_data)

evaluator_llm = LangchainLLMWrapper(OllamaLLM(model="gemma3:27b", base_url="http://localhost:11434"))
metric = OpenFactualCorrectness(llm=evaluator_llm)
score = metric.single_turn_score(sample)

print(score)  # e.g. 1.0

☁️ Hosted LLM (e.g., OpenAI)

from openai import OpenAI
from ragas import SingleTurnSample
from ragas.llms import LangchainLLMWrapper

from truthscore import OpenFactualCorrectness

evaluator_llm = LangchainLLMWrapper(OpenAI())
metric = OpenFactualCorrectness(llm=evaluator_llm)

# test_data same as above
score = metric.single_turn_score(SingleTurnSample(**test_data))

⚙️ Custom NLI Models

import torch
from langchain_community.llms import OllamaLLM
from ragas import SingleTurnSample
from ragas.llms import LangchainLLMWrapper
from sentence_transformers import CrossEncoder

from truthscore import OpenFactualCorrectness

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
nli_model = CrossEncoder("cross-encoder/nli-deberta-v3-large")
nli_model.model.to(device)

evaluator_llm = LangchainLLMWrapper(OllamaLLM(model="gemma3:27b", base_url="http://localhost:11434"))
metric = OpenFactualCorrectness(llm=evaluator_llm, nli_model=nli_model)

# test_data same as above
score = metric.single_turn_score(SingleTurnSample(**test_data))

📊 Background

This metric was evaluated across a 500-example benchmark using perturbation levels A0–A4 on top of the Google Natural Questions dataset using truthbench.

See full results in the trutheval project.