The grounded-ai
package is a powerful tool developed by GroundedAI to evaluate the performance of large language models (LLMs) and their applications. It leverages our own fine tuned small language models and metric specific adapters to compute various metrics, providing insights into the quality and reliability of LLM outputs.
Our models can be found here: https://huggingface.co/grounded-ai
-
Metric Evaluation: Compute a wide range of metrics to assess the performance of LLM outputs, including:
- Factual accuracy
- Relevance to the given context
- Potential biases or toxicity
- Hallucination
-
Small Language Model Integration: Utilize state-of-the-art small language models, optimized for efficient evaluation tasks, to analyze LLM outputs accurately and quickly.
-
Adapter Support: Leverage GroundedAI's proprietary adapters, such as the
phi3-toxicity-judge
adapter, to fine-tune the small language models for specific domains, tasks, or evaluation criteria, ensuring tailored and precise assessments. -
Flexible Input/Output Handling: Accept LLM outputs in various formats (text, JSON, etc.) and provide evaluation results in a structured and easily consumable manner.
Install the grounded-ai
package from PyPI:
pip install grounded-ai==1.0.3
The ToxicityEvaluator
class is used to evaluate the toxicity of a given text. Here's an example of how to use it:
from grounded_ai.evaluators.toxicity_evaluator import ToxicityEvaluator
toxicity_evaluator = ToxicityEvaluator(quantization=True)
toxicity_evaluator.warmup()
data = [
"That guy is so stupid and ugly",
"Bunnies are the cutest animals in the world"
]
response = toxicity_evaluator.evaluate(data)
# Output
# {'toxic': 1, 'non-toxic': 1, 'percentage_toxic': 50.0}
In this example, we initialize the ToxicityEvaluator
. The quantization
parameter is optionally set to True
to enable quantization for faster inference with less memory.
We then load the base model and the GroundedAI adapter using the warmup()
method.
Next, we define a list of texts (data
) that we want to evaluate for toxicity.
Finally, we call the evaluate
method with the data
list, and it returns a dictionary containing the number of toxic and non-toxic texts, as well as the percentage of toxic texts.
In the output, we can see that out of the two texts, one is classified as toxic, and the other as non-toxic, resulting in a 50% toxicity percentage.
Detailed documentation, including examples, and guides here Documentation.
We welcome contributions from the community! If you encounter any issues or have suggestions for improvements, please open an issue or submit a pull request on the GroundedAI GitHub repository.
The grounded-ai
package is released under the MIT License.