Pinned Loading

1. kimi-k3-from-scratch-kda-attention-residuals-and-stable-latentmoe kimi-k3-from-scratch-kda-attention-residuals-and-stable-latentmoe Public

   Build every architectural innovation from the Kimi K3 technical report in numpy at toy scale: Kimi Delta Attention with lower-bounded decay and an exact chunkwise-parallel form, Gated MLA with NoPE…

   Python 2

2. tiny-gpt-from-scratch tiny-gpt-from-scratch Public

   Build a small character-level GPT end-to-end in pure NumPy, starting from tokenization and array basics and ending with multi-head self-attention, Adam, and sampling. Each step grows the same codeb…

   Python 1

3. multimodal-autoregressive-image-generator-from-scratch-in-jax multimodal-autoregressive-image-generator-from-scratch-in-jax Public

   Build a text-conditioned image generator end to end in JAX by training a VQ-VAE to turn tiny images into discrete tokens, then an autoregressive transformer that models text-prefixed token sequence…

   Python 1

4. rectified-flow-from-scratch rectified-flow-from-scratch Public

   Build a complete rectified-flow / linear interpolant flow-matching generative model in pure PyTorch. Instead of a long discrete reverse chain that predicts noise, you learn a velocity field that tr…

   Python 1

5. reinforcement-learning-for-tic-tac-toe-from-minimax-to-dqn reinforcement-learning-for-tic-tac-toe-from-minimax-to-dqn Public

   Build a complete Tic-Tac-Toe reinforcement learning lab from scratch: the game engine, minimax baselines, tabular Q-learning with self-play, and a deep Q-network agent, then compare value-based and…

   Python 1

6. verillm-publicly-verifiable-decentralized-llm-inference-from-scratch-in-pytorch verillm-publicly-verifiable-decentralized-llm-inference-from-scratch-in-pytorch Public

   Build a minimal GPT-style transformer with a KV cache from scratch, then wrap its inference in a Merkle-commitment and spot-check protocol that lets untrusted workers serve LLM outputs verifiably. …

   Python 1