vid1203/festo-llm-experiment-b

Experiment B — Receding Horizon MPC, 9,000 Training Records

Paper: LLMs as Autonomous Controllers for Multivariable Industrial Fluid Processes Author: Vidyashree Rayar — BTU Cottbus-Senftenberg

Introduces 3-step predictive horizon MPC with level-biased cost function (2×level + 1×temp). SP+CoT+P achieves 100% level accuracy. Content hallucination rises to 39.2%, making the safety layer essential.

Base Model

meta-llama/Llama-3.2-3B-Instruct fine-tuned with QLoRA (4-bit NF4, LoRA r=16, alpha=32, rsLoRA enabled).

Code & Results

github.com/vidyashreerayar/festo-llm-process-control