Myst-Super-Coder-75M

A 75M-parameter multi-language instruction-tuned decoder-only transformer trained from scratch.

Architecture

• Parameters: ~75M
• Layers: 12
• Attention Heads: 12
• Embedding Dimension: 768
• Context Length: 1024 tokens
• Format: ChatML (<|im_start|>user, <|im_start|>assistant)

Training & Licensing

• Weights & Architecture: MIT License
• Training Data: Open-source CodeAlpaca instruction dataset & permissive multi-language code.