Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.murnitur.ai/llms.txt

Use this file to discover all available pages before exploring further.

Murnitur AI allows developers to run programmatic evaluations, providing a flexible and automated way to assess Large Language Models (LLMs) through code. Note: Before running programmatic evaluations, you need to set up the Murnitur SDK with your API key. Detailed instructions on how to set up the SDK can be found here.

​
Example


import murnitur
from murnitur.loaders import Loader
from murnitur.annotation.eval import Evaluation

# initialize evaluation
murnitur.set_api_key("mt-ey...")
key = "sk-..."

eval = Evaluation(openai_key=key, murnitur_key=murnitur.get_api_key())

# Load datasets from csv
data = Loader.load_csv("https://murnitur.github.io/murnix-kube/sample-dataset.csv")

results = eval.run_suite(metrics=["hallucination", "pii"], dataset=data)

print(results)
Each run is saved in the UI. You can view the results on the AI Evaluations page.

​
Knowledge Retention Evaluation

To evaluate how well your model retains information across interactions, use Murnitur’s knowledge_retention method. You can also set up a base threshold to define acceptable retention levels.
  1. Import the Required Module: 
    from murnitur.annotation.eval import Evaluation
  2. Prepare Your Dataset: 
    data = [
  {
    "input": "How often should I run webinars to gain traction for Murnitur?",
    "output": "Once a month to start. This gives you time to prepare quality content."
  },
  {
    "input": "What topics should I cover in these webinars?",
    "output": "Focus on unique features like tracing and monitoring, and AI safety with Murnitur Shield."
  }
]
  3. Set Up a Base Threshold (Optional): 
    base_threshold = 0.2  # Define your acceptable retention level
  4. Run the Evaluation: 
    eval = Evaluation()
results = eval.knowledge_retention(dataset=data, save_output=True, threshold=base_threshold)
print(results)
  5. Sample Output: 
    {
  "score": 0.1,
  "reason": "The model retains information well with minimal loss."
}
This method returns a JSON with a score (0-1) indicating the level of knowledge retention, where closer to 0 means better retention. By setting a base threshold, you can determine if the retention meets your desired standards.

​
Creating dataset

Side Note: If you are creating a dataset from a dictionary, use the following data structure: 
dataset = [
    {
        "input": "",
        "output": "",
        "ground_truth": "",
        "context": [""],
        "retrieval_context": [""],
    }
]

​
Loading dataset

Use any of the following functions to load the dataset: 
from murnitur.loaders import Loader

dataset_from_json = Loader.load_json(path_to_json_file)

dataset_from_csv = Loader.load_csv(path_to_csv_document)

​
Extra Params

Extra parameters that can be passed to run_suite
  • save_output: A boolean. Set to false if you do not want to store the results in Murnitur AI.
  • async_mode: Indicates whether the requests should run asynchronously.

​
Evaluation Metrics

We currently support the following evaluation metrics:
  • hallucination: Measures the accuracy of the model’s output, identifying any fabricated or incorrect information.
  • faithfulness: Evaluates how well the model’s output adheres to the provided context or source material.
  • relevancy: Assesses the relevance of the model’s response to the input query or prompt.
  • bias: Detects any unfair or discriminatory tendencies in the model’s predictions.
  • context-relevancy: Determines how relevant the input or output is in relation to the given context.
  • context-precision: Measures the precision of the model’s output within the specific context provided.
  • toxicity: Identifies harmful or offensive language in the model’s responses.
  • summarization: Evaluates the quality and accuracy of summaries generated by the model.
  • pii: Detects the presence of personally identifiable information in the model’s output.

​
Metrics & Required Fields

MetricPayload
hallucinationinput, output, context
faithfulnessinput, output, retrieval_context
relevancyinput, output
biasinput, output
context-relevancyinput, output, retrieval_context
context-precisioninput, output, ground_truth, retrieval_context
toxicityinput, output
summarizationinput, output
piioutput use murnitur-shield to detect from the input