metadata
license: mit
language:
- en
- ru
tags:
- rust
- custom-architecture
- text-generation
DeepCode-Rust-2M
This is a lightweight, custom language model built entirely from scratch in Rust. It does not rely on Python, PyTorch, or standard Hugging Face transformers libraries.
Model Details
- Parameters: ~2 Million
- Architecture: Decoder-only Transformer (6 layers, 256 embedding dim)
- Format: Serialized via
bincode(requires the custom Rust inference engine)
Benchmark Results (Napkin Benchmark)
Tested on 6 basic prompts:
- Python Addition: 4/5 (Successfully generated
def add(a, b): return a + b) - Python Multiplication: 2/5 (Recognized
multiply, but broken syntax) - Python Loop: 0/5 (Stuck in a repetition loop)
- HTML Structure: 1/5 (Recognized tags, but no hierarchical structure)
- HTML Button: 1/5
- Overall Score: ~2 out of 5
Limitations
Due to its extremely small size, the model often suffers from repetition loops and fails to understand nested structures (like HTML trees). It is an educational project demonstrating how language models work under the hood.