Autonym

Automated ML Run Naming & Summarization using LLMs.

This is a vibe-coded tool to help me track my protoyping experiments. Autonym experiment scribe reads your current git diff, separates logic changes from config updates, and uses an LLM (Local or Remote) to generate a semantic run_name and technical description.

It solves the problem of manual run tracking ("run-42", "run-final-final") by ensuring every experiment has a descriptive, auto-generated label based on what changed in the code.

Features

• Smart Diffing: Separates code logic from YAML config changes.
• Grammar Forcing: Uses Pydantic grammar forcing to output valid JSON to avoid excessively long responses.
• Provider Agnostic: Supports local Ollama (free, private) and OpenAI (remote, fast).

Usage

import wandb
from autonym import Autonym

# ... config setup ...

scribe = Autonym(provider="ollama", model_name="phi3.5")

# Generate run metadata from git diffs & config changes
meta = scribe.summarize_run(runtime_config=config, ref_config_path="config.yaml", base_ref="origin/main")

if meta:
    wandb.init(name=meta.run_name, notes=meta.description, config=config)
else:
    wandb.init(config=config)

# ... training loop ...
train(config)

Installation

# 1. Install Python dependencies
pip install ollama openai pydantic

# 2. (Optional) For local inference, install Ollama
# [https://ollama.com/download](https://ollama.com/download)
ollama pull phi3.5