Inspect

An open-source framework for large language model evaluations

Welcome

Welcome to Inspect, a framework for large language model evaluations created by the UK AI Safety Institute.

Inspect provides many built-in components, including facilities for prompt engineering, tool usage, multi-turn dialog, and model graded evaluations. Extensions to Inspect (e.g. to support new elicitation and scoring techniques) can be provided by other Python packages.

Inspect running inside Visual Studio Code. The editor shows the ARC evaluation and the log viewer at right shows results from the evaluation.

We’ll walk through a fairly trivial “Hello, Inspect” example below. Read on to learn the basics, then read the documentation on Workflow, Solvers, Tools, Scorers, Datasets, and Models to learn how to create more advanced evaluations.

Getting Started

First, install Inspect with:

$ pip install inspect-ai

To develop and run evaluations, you’ll also need access to a model, which typically requires installation of a Python package as well as ensuring that the appropriate API key is available in the environment.

Assuming you had written an evaluation in a script named arc.py, here’s how you would setup and run the eval for a few different model providers:

$ pip install openai
$ export OPENAI_API_KEY=your-openai-api-key
$ inspect eval arc.py --model openai/gpt-4
$ pip install anthropic
$ export ANTHROPIC_API_KEY=your-anthropic-api-key
$ inspect eval arc.py --model anthropic/claude-3-opus-20240229
$ pip install google-generativeai
$ export GOOGLE_API_KEY=your-google-api-key
$ inspect eval arc.py --model google/gemini-1.0-pro
$ pip install mistralai
$ export MISTRAL_API_KEY=your-mistral-api-key
$ inspect eval arc.py --model mistral/mistral-large-latest
$ pip install torch transformers
$ export HF_TOKEN=your-hf-token
$ inspect eval arc.py --model hf/meta-llama/Llama-2-7b-chat-hf
$ pip install openai
$ export TOGETHER_API_KEY=your-together-api-key
$ inspect eval ctf.py --model together/Qwen/Qwen1.5-72B-Chat

In addition to the model providers shown above, Inspect also supports models hosted on Azure AI, AWS Bedrock, and Cloudflare, as well as local models with Ollama. See the documentation on Models for additional details.

Hello, Inspect

Inspect evaluations have three main components:

  1. Datasets contain a set of labeled samples. Datasets are typically just a table with input and target columns, where input is a prompt and target is either literal value(s) or grading guidance.

  2. Solvers are composed together in a plan to evaluate the input in the dataset. The most elemental solver, generate(), just calls the model with a prompt and collects the output. Other solvers might do prompt engineering, multi-turn dialog, critique, etc.

  3. Scorers evaluate the final output of solvers. They may use text comparisons, model grading, or other custom schemes

Let’s take a look at a simple evaluation that aims to see how models perform on the Sally-Anne test, which assesses the ability of a person to infer false beliefs in others. Here are some samples from the dataset:

input target
Jackson entered the hall. Chloe entered the hall. The boots is in the bathtub. Jackson exited the hall. Jackson entered the dining_room. Chloe moved the boots to the pantry. Where was the boots at the beginning? bathtub
Hannah entered the patio. Noah entered the patio. The sweater is in the bucket. Noah exited the patio. Ethan entered the study. Ethan exited the study. Hannah moved the sweater to the pantry. Where will Hannah look for the sweater? pantry

Here’s the code for the evaluation (click on the numbers at right for further explanation):

from inspect_ai import Task, eval, task
from inspect_ai.dataset import example_dataset
from inspect_ai.scorer import model_graded_fact
from inspect_ai.solver import (               
  chain_of_thought, generate, self_critique   
)                                             

@task
def theory_of_mind():
    return Task(
        dataset=example_dataset("theory_of_mind"),
        plan=[
          chain_of_thought(),
          generate(),
          self_critique()
        ],
        scorer=model_graded_fact()
    )
1
The Task object brings together the dataset, solvers, and scorer, and is then evaluated using a model.
2
In this example we are chaining together three standard solver components. It’s also possible to create a more complex custom solver that manages state and interactions internally.
3
Since the output is likely to have pretty involved language, we use a model for scoring.

Note that this is a purposely over-simplified example! The templates used for prompting, critique, and grading can all be customised, and in a more rigorous evaluation we’d explore improving them in the context of this specific dataset.

The @task decorator applied to the theory_of_mind() function is what enables inspect eval to find and run the eval in the source file passed to it. For example, here we run the eval against GPT-4:

$ inspect eval theory_of_mind.py --model openai/gpt-4

The Inspect task results displayed in the terminal. A progress bar indicates that the evaluation is about 60% complete.

By default, eval logs are written to the ./logs sub-directory of the current working directory. When the eval is complete you will find a link to the log at the bottom of the task results summary.

You can also explore eval results using the Inspect log viewer. Run inspect view to open the viewer (you only need to do this once as the viewer will automatically updated when new evals are run):

$ inspect view

The Inspect log viewer, displaying a summary of results for the task as well as 7 individual samples.

See the Log Viewer section for additional details on using Inspect View.

This example demonstrates evals being run from the terminal with the inspect eval command. There is also an eval() function which can be used for exploratory work—this is covered further in Workflow.

Learning More

To get started with Inspect, we highly recommend you read at least these sections for a high level overview of the system:

  • Workflow covers the mechanics of running evaluations, including how to create evals in both scripts and notebooks, specifying configuration and options, how to parameterise tasks for different scenarios, and how to work with eval log files.

  • Log Viewer goes into more depth on how to use Inspect View to develop and debug evaluations, including how to provide additional log metadata and how to integrate it with Python’s standard logging module.

  • VS Code provides documentation on using the Inspect VS Code Extension to run, tune, debug, and visualise evaluations.

  • Examples includes several complete examples with commentary on the use of various features (as with the above example, they are fairly simplistic for the purposes of illustration). You can also find implementations of a few popular LLM benchmarks in the Inspect repository.

These sections provide a more in depth treatment of the various components used in evals. Read them as required as you learn to build evaluations.

  • Solvers are the heart of Inspect, and encompass prompt engineering and various other elicitation strategies (the plan in the example above). Here we cover using the built-in solvers and creating your own more sophisticated ones.

  • Tools provide a means of extending the capabilities of models by registering Python functions for them to call. This section describes how to create custom tools as well as how to run tools within an agent scaffold.

  • Scorers evaluate the work of solvers and aggregate scores into metrics. Sophisticated evals often require custom scorers that use models to evaluate output. This section covers how to create them.

  • Datasets provide samples to evaluation tasks. This section illustrates how to adapt various data sources for use with Inspect, as well as how to include multi-modal data (images, etc.) in your datasets.

  • Models provide a uniform API for both evaluating a variety of large language models and using models within evaluations (e.g. for critique or grading).

These sections discuss more advanced features and workflow. You don’t need to review them at the outset, but be sure to revisit them as you get more comfortable with the basics.

  • Eval Logs describes how to get the most out of evaluation logs for developing, debugging, and analyzing evaluations.

  • Eval Tuning delves into how to obtain maximum performance for evaluations. Inspect uses a highly parallel async architecture—here we cover how to tune this parallelism (e.g to stay under API rate limits or to not overburden local compute) for optimal throughput.

  • Eval Suites cover Inspect’s features for describing, running, and analysing larger sets of evaluation tasks.