byte-level

A tiny byte-level multi-head content classifier (~1.5M params, ~200KB ONNX, <6ms). Classifies code, text, markup, config, images, binary, secrets, 62 code languages, 30 text languages, 90 MIME types from raw bytes — no tokenizer needed.

An editable, auditable 807K-param byte-level LLM: CRUD single facts with provable per-edit locality, and abstain when unsure instead of guessing. CPU, offline.

Kronecker Embeddings: byte-level structured token representations for parameter-efficient language models. Reference implementation.

🐍This is a fast, lightweight, and clean CPython extension for the Byte Pair Encoding (BPE) algorithm, which is commonly used in LLM tokenization and NLP tasks.

An efficient openblocks parser module.

Tiny byte-level AI on CPU: a zero-dependency C++17 runtime and neural decision models (0.8 ms per 64-byte decision, 9.4 MB RAM). Reads raw bytes, no tokenizer. Specialists for secret scanning (F1 0.797 vs 0.337 for gitleaks on CredData) and PII redaction. Code under Apache-2.0; model weights under their own license.

A deterministic byte-level BPE tokenizer in pure Python, built from scratch with strict tests, typed code, and polished docs.

Byte-level GPT from scratch in pure PyTorch: single-file training loop, 26M C# boilerplate generator & 38M toy Turkish chat model.

Open HSL/HoLo research: a deterministic byte-native signal substrate tested across language, speech, and diffusion with matched controls and honest limits.

Four-seed matched ASR feasibility study using an HSL byte-signal substrate and spectral lens. Open code, weights, controls, and limitations; not SOTA.

Hierarchical ternary byte-level language models — benchmark harness

File security system using remote authentication

NeuralPiece: An adaptive byte-level neural tokenizer designed to surpass traditional BPE and Unigram via deep learning chunking.

Engram without a tokenizer

A tokenizer-free byte-level LM in MLX: entropy-based dynamic patching around a local encoder/decoder and a latent global transformer over patches.

Tokenizer-free HSL text-to-speech feasibility prototype with open code and weights. Working synthesis evidence; formal perceptual evaluation remains open.

A tiny byte-level neural network, trainable from scratch using only the Python standard library. Runs on Termux and Android.

PyTorch transformer input from raw bytes with zero learned input parameters: no tokenizer, embedding table, or learned input projection.

Some investigations I made with AMD cloud credits: training language models from scratch, exploring low-bit and byte-level architectures, and running ablation studies.

Language-model proof of operation for HSL's zero-learned-parameter input path. Open code, weights, and documented limitations; not a SOTA claim.