Custom Bakabaka Language LLM

I need a specialist who can take the idea of a Large Language Model trained to speak, understand, and generate the ancient “bakabaka” tongue and turn it into a working system. The top priority, as clarified in my brief, is Custom LLM development; all other tasks orbit around that central goal.

Scope
• Build a transformer-based architecture that follows my in-house “nitro_spectro” design. The design targets a 0 % compression rate, so every token and weight must remain lossless throughout preprocessing, training, and export.
• Train the model end-to-end so it fluently produces bakabaka text while also handling input prompts in today’s commonly spoken languages (English, Spanish, Mandarin, etc.) and returning answers in bakabaka.
• Provide a clean inference pipeline (Python preferred, PyTorch or TensorFlow acceptable) that I can host on-prem or on a GPU cloud instance.

Data
I will supply a curated parallel corpus: modern-language prompts aligned with bakabaka references. You are welcome to augment it with publicly available multilingual datasets to strengthen the encoder side, but the bakabaka output quality must stay paramount.

Acceptance Criteria
1. The model achieves at least 85 % BLEU on a held-out bakabaka validation set I will share.
2. It answers modern-language prompts in bakabaka with