05 / Credits

A new identity, with its technical roots left visible.

Jarvisn't is Temuujin's project. It also stands on thoughtful open-source work that deserves direct credit.

Technical foundation

nanochat

Jarvisn't is built on top of nanochat, an open-source language-model training project by Andrej Karpathy. Its compact end-to-end training stack remains the technical foundation of this repository.

Visit upstream repository ↗

Ideas and experimentation

modded-nanogpt

I also looked at modded-nanogpt for ideas and inspiration while building Jarvisn't, especially around efficient model training and optimization.

Explore modded-nanogpt ↗

PyTorch

Tensor operations, model definition, optimization, and distributed execution. See the official PyTorch documentation ↗.

RustBPE

The tokenizer implementation used by the training pipeline.

FlashAttention

Hardware-aware attention acceleration, with a PyTorch SDPA fallback.

Next.js

The framework behind this documentation website.

RunPod

Serverless infrastructure hosting CPU inference for the chat service. The worker loads the trained model checkpoint on demand, so the service can scale down when idle.

RunPod Serverless ↗

Implementation citations, dataset-source URLs, and historical comments remain in the code where they provide useful provenance.