modsys: model management tool

Framework for improving AI models

Docs»

Join the waitlist · Report Bug · Community Discord

ModsysML is an open source toolbox for automating, building insights and evaluating model responses across relevant testcases. Prompt tuning is a way to fine-tune AI models with a small context window of data. Instead of foundational training, prompt tuning allows users to take pre-trained models and tweak them to a specific use-case. Instead of stopping here we've incorporated the ability to automate evaluations and even suggestions for model responses while giving users the ability to connect to any database/api resource for curating analytics alongside running evaluations.

ModsysML provides a framework for building relevant test cases across 4 domains, semantic, LLM based, Human in the loop and programmatic tests. It's a collection of python plugins wrapped into an SDK that provide the basic functionality, but may not fit each use case out of the box. You can craft custom test cases, run evaluations on model responses, curate automated scripts and generate data frames.

ModsysML strictly uses official provider integrations via API, psycopg and a few other modules to provide the building blocks. ModsysML is the upstream project to the Apollo API interface which serves as a UI for your development needs.

🚀 Interesting, how can I try it?

Lets install the SDK first...

pip install modsys

Regression tests vs Automated pipelines

modsys helps you tune LLM prompts systematically across many relevant test cases. By evaluating and comparing LLM outputs to build decision making workflows. Users can test prompt quality and catch regressions faster.

Evaluating prompt quality

With Modsys python library and CLI toolkit, you can:

Detecting real-time changes in your data
Automating tasks against image, video, audio or text
Simplifying the process of back-testing quality for your AI models
Making sure your integration is robust, so you never again have to worry about stuck/stale data or false-positives
Test multiple prompts against predefined test cases
Evaluate quality and catch regressions by comparing LLM outputs side-by-side
Speed up evaluations with caching and concurrent tests
Flag bad outputs automatically by setting "expectations"
Use as a command line tool, or integrate into your workflow with our library
Use any AI provider, API or database under one API

modsys produces table views that allow you to quickly review prompt outputs across many inputs. The goal: tune prompts systematically across all relevant test cases, instead of testing prompts by trial and error.

Usage (command line)

Support for user interface coming soon

It works on the command line, you can output to [json, csv, yaml]:

To get started, run the following command:

modsys init

This will create some templates in your current directory: prompts.txt, vars.csv, and config.json.

After editing the prompts and variables to your desired state, modsys command to kick off an prompt evaluation test:

modsys -p ./prompts.txt -v ./vars.csv -r openai:completion

If you're looking to customize your usage, you have a wide set of parameters at your disposal. See the Configuration docs for more detail:

Option	Description
`-p, --prompts <paths...>`	Paths to prompt files, directory, or glob
`-r, --providers <name or path...>`	One of: openai:chat, openai:completion, openai:model-name, hive:hate, google:safety, etc. See AI Providers
`-o, --output <path>`	Path to output file (csv, json, yaml, html)
`-v, --vars <path>`	Path to file with prompt variables (csv, json, yaml)
`-c, --config <path>`	Path to configuration file. `config.json` is automatically loaded if present
`-j, --max-concurrency <number>` coming soon	Maximum number of concurrent API calls
`--table-cell-max-length <number>` coming soon	Truncate console table cells to this length
`--grader` coming soon	Provider that will grade outputs, if you are using

Examples

Prompt quality

In this example, we evaluate whether adding adjectives to the personality of an chat bot affects the responses:

modsys -p prompts.txt -v vars.csv -r openai:gpt-3.5-turbo

This command will evaluate the prompts in prompts.txt, substituing the variable values from vars.csv, and output results in your terminal.

Have a look at the setup and full output in another format:

modsys -p prompts.txt -v vars.csv -r openai:gpt-3.5-turbo -o output.json

You can also output a nice spreadsheet, JSON, or YAML file:

{
  "results": [
    {
      "prompt": {
        "raw": "Rephrase this in French: Hello world",
        "display": "Rephrase this in French: {{body}}"
      },
      "vars": {
        "body": "Hello world"
      },
      "response": {
        "output": "Bonjour le monde",
        "tokenUsage": {
          "total": 19,
          "prompt": 16,
          "completion": 3
        }
      }
    }
    // ...
  ],
  "stats": {
    "successes": 4,
    "failures": 0,
    "tokenUsage": {
      "total": 120,
      "prompt": 72,
      "completion": 48
    }
  }
}

Here's an example of a side-by-side comparison of multiple prompts and inputs:

Model quality

You can evaluate the difference between safety outputs for a specific context:

Model quality tests & python package for model testing is a beta feature at the moment, open an issue and tag us to setup

modsys -p prompts.txt -r hiveai:hate google:safety -o output.json

Configuration

Setting up an model test: Learn more about how to set up prompt files, vars file, output, etc.

Building Automated Pipelines in the User Interface or Programmatically

Let's setup your first Integration!

It will pull from your local database (and keep it in sync).

# import the package
from modsys.client import Modsys

# sync data from your database instance
# (we support supabase at the current moment or postgresql via uri format)
Modsys.connect("postgres://username:password@hostname:port/database_name")

# If you want to test out operation on your external connection
Modsys.fetch_tables()
Modsys.query("desc", "table", "column")

...and create a workflow with a simple command:

# import the package
from modsys.client import Modsys

# Use any provider
Modsys.use("google_perspective:<model name>", google_perspective_api_key="YOUR_API_TOKEN_HERE")

# An option for image detection, connect to sightengine provider or other image service first
Modsys.detectImage('https://example.com/some-endpoint') # Image Analysis/OCR

# Lets check to see if a phrase contains threats
Modsys.detectText(prompt="Phrase1", content_id="content-id", community_id="user-id")

Example response:

{
  "attributeScores": {
    "THREAT": {
      "spanScores": [
        {
          "begin": 0,
          "end": 12,
          "score": { "value": 0.008090926, "type": "PROBABILITY" }
        }
      ],
      "summaryScore": { "value": 0.008090926, "type": "PROBABILITY" }
    },
    "INSULT": {
      "spanScores": [
        {
          "begin": 0,
          "end": 12,
          "score": { "value": 0.008804884, "type": "PROBABILITY" }
        }
      ],
      "summaryScore": { "value": 0.008804884, "type": "PROBABILITY" }
    },
    "SPAM" // ...
  },
  "languages": ["en"],
  "clientToken": "content_123",
  "detectedLanguages": ["en", "fil"]
}

Experimental inputs:

# Create custom rules which creates a task!
Modsys.rule('Phrase1', '>=', '0.8')

Modsys.detectImage('Image1', 'contains', 'VERY_LIKELY') # Image Analysis/OCR
Modsys.detectSpeech('Audio1', 'contains', 'UNLIKELY') # Audio Processing
Modsys.detectVideo('Video1', 'contains', 'POSSIBLE') # Video Analysis
Modsys.detectText('Phrase1', 'contains', 'UNKNOWN') # Text Analysis
Modsys.test('prompt', 'expected_output') # ML Validation

That's all it takes!

In practice, you probably want to use one of our native SDKs to interact with different AI providers or use our custom browser client so you dont have to write code. If so, sign up for the downstream Apollo ModsysML Console!

Cool, what can I build with it?

Modsys can help you quickly automate tasks for model management, performance, labeling, object detection and more.
Teams can use Modsys to build native in-app connections related to active response, content moderation, risk management, fraud detection, etc.
Some automate their personal lives with Modsys by integrating against discord communities or their personal lives

Development

Contributions are welcome! Please feel free to submit a pull request or open an issue.

📦 pre-commit config

As an open source project, Modsys welcomes contributions from the community at large. This isn’t an exhaustive reference and is a living document subject to change as needed when the project formalizes any practice or pattern.

Clone the repo and start Modsys locally...

git clone https://github.com/modsysML/modsys.git
cd modsys && python3 -m venv env && source env/bin/activate && pip install -r requirements.txt

After installing system dependencies be sure to install pre-commit for lint checks

pip install pre-commit

pre-commit install

pre-commit run --all-files

Modsys uses commit messages for automated generation of project changelog. For every pull request we request contributors to be compliant with the following commit message notation.

<type>: <summary>

<body>

Accepted <type> values:

new = newly implemented user-facing features
chg = changes in existing user-facing features
fix = user-facing bugfixes
oth = other changes which users should know about
dev = any developer-facing changes, regardless of new/chg/fix status

Summary (The first line)

The first line should not be longer than 75 characters, the second line is always blank and other lines should be wrapped at 80 characters.

🔍 Neat, I would like to learn more

⭐ Follow our development by starring us here on GitHub ⭐

Share feedback or chat with us on the Discord community

AI Providers

We support OpenAI as well as a number of models. It's also possible to set up your own custom AI provider.

apolloapi/modsysML