ravipatib / mercylm

What is mercylm?

MercyLM lets you train a 15M parameter character LLM from scratch in Python using PyTorch, embodying Mercy—a network engineer surviving alone in an empty world with only a pigeon named Gerald for company. Run simple CLI commands to prepare synthetic conversations, train the model in under two hours on a local Mac, and chat interactively via terminal or export to Hugging Face. It's a full pipeline for tiny, focused language models without cloud dependency.

Why is it gaining traction?

Unlike massive pretrained models, MercyLM trains entirely locally on consumer hardware like Apple Silicon, turning hardware limits into character traits—short memory and simple responses fit her isolated life perfectly. It resists jailbreak attempts out of the box, staying in character, and includes eval cases to test coherence on held-out prompts. Developers dig the end-to-end flow from data gen to inference, all in one repo.

Who should use this?

ML hobbyists building their first from-scratch LLM to grasp training loops without GPU farms. Indie game devs crafting dialogue for survival characters in empty worlds. Python scripters experimenting with character LLMs for chatbots or story prototypes on laptops.

Verdict

Grab it if you want a quick, local LLM project—docs are solid, CLI is intuitive, but with 12 stars and 1.0% credibility score, treat it as an educational prototype, not production-ready. Fork and tweak for your own 15M parameter experiments.

(187 words)