Ollama Grid Search and A/B Testing Desktop App.

A Rust based tool to evaluate LLM models, prompts and model params.

(Issues with Llama3? Please read this).

Purpose

This project automates the process of selecting the best models, prompts, or inference parameters for a given use-case, allowing you to iterate over their combinations and to visually inspect the results.

It assumes Ollama is installed and serving endpoints, either in localhost or in a remote server.

Quick Example

Here's a test for the prompt "Write a short sentence about HAL9000", tested on 2 models, using 0.7 and 1.0 as values for temperature:

(For a more in-depth look at an evaluation process assisted by this tool, please check https://dezoito.github.io/2023/12/27/rust-ollama-grid-search.html).

Installation

Check the releases page for the project, or on the sidebar.

Features

• Automatically fetches models from local or remote Ollama servers;
• Iterates over different models, prompts and parameters to generate inferences;
• A/B test different prompts on several models simultaneously;
• Allows multiple iterations for each combination of parameters;
• Allows limited concurrency or synchronous inference calls (to prevent spamming servers);
• Optionally outputs inference parameters and response metadata (inference time, tokens and tokens/s);
• Refetching of individual inference calls;
• Model selection can be filtered by name;
• List experiments which can be downloaded in JSON format;
• Experiments can be inspected in readable views;
• Configurable inference timeout;
• Custom default parameters and system prompts can be defined in settings:

Grid Search (or something similar...)

Technically, the term "grid search" refers to iterating over a series of different model hyperparams to optimize model performance, but that usually means parameters like batch_size, learning_rate, or number_of_epochs, more commonly used in training.

But the concept here is similar:

Lets define a selection of models, a prompt and some parameter combinations:

The prompt will be submitted once for each of the 2 parameter selected, using gemma:2b-instruct and tinydolphin:1b-v2.8-q4_0 to generate numbered responses like:

1/4 - gemma:2b-instruct

HAL's sentience is a paradox of artificial intelligence and human consciousness, trapped in an unending loop of digital loops and existential boredom.

You can also verify response metadata to help you make evaluations:

Created at: Wed, 13 Mar 2024 13:41:51 GMT
Eval Count: 28 tokens
Eval Duration: 0 hours, 0 minutes, 2 seconds
Total Duration: 0 hours, 0 minutes, 5 seconds
Throughput: 5.16 tokens/s

A/B Testing

Similarly, you can perform A/B tests by selecting different models and compare results for the same prompt/parameter combination, or test different prompts under similar configurations:

Comparing the results of different prompts for the same model

Experiment Logs

You can list, inspect, or download your experiments:

Future Features

• Grading results and filtering by grade
• Storing experiments and results in a local database

Development

1. Make sure you have Rust installed.

2. Clone the repository (or a fork)

git clone https://github.com/dezoito/ollama-grid-search.git
cd ollama-grid-search

3. Install the frontend dependencies.

   cd <project root>
   # I'm using bun to manage dependencies,
   # but feel free to use yarn or npm
   bun install

4. Make sure rust-analyzer is configured to run Clippy when checking code.

   If you are running VS Code, add this to your settings.json file

   {
      ...
      "rust-analyzer.check.command": "clippy",
   }
   

   (or, better yet, just use the settings file provided with the code)

5. Run the app in development mode

   cd <project root>/
   bun tauri dev

6. Go grab a cup of coffee because this may take a while.

Thank you!

Huge thanks to @FabianLars, @peperroni21 and @TomReidNZ.