LightfieldAI-native CRM that builds itself and does work for you
Promoted
Launching today
Attenta
A from-scratch Rust Transformer trained on legal documents
2 followers
A from-scratch Rust Transformer trained on legal documents
2 followers
Attenta is a from-scratch Rust implementation of the Transformer architecture (Vaswani et al. 2017) built without PyTorch or TensorFlow. Designed for high efficiency, it features multi-head attention, BPE tokenization, and custom matrix multiplication. This release comes pre-trained on Indian legal documents, showcasing a pure Rust pipeline for specialized NLP tasks from tokenization to beam search decoding. Fully tested with 85 unit tests and ready to self-host.
Hey Product Hunt community! š
I'm the creator of Attenta. I wanted to see how far I could push pure Rust for deep learning architectures without relying on massive frameworks like PyTorch or TensorFlow.
Attenta is a complete, from-scratch implementation of the original "Attention Is All You Need" Transformer paper. To make it a true end-to-end proof of concept, I trained it specifically on structured Indian legal documents.
Key features engineered directly into the crate:
ā¢ Custom BPE tokenization & bucket batching
ā¢ Multi-head attention with optimized 64x64 blocked matmul for cache efficiency
ā¢ Adam optimizer with warmup schedules & label smoothing
ā¢ 85 unit tests covering everything from scratch
Whether you love Rust, are curious about zero-dependency AI, or work in legal tech, I'd love for you to check out the repository, run the training scripts, and let me know your thoughts!
What features or optimization techniques should I implement next? Let me know below! š