| --- |
| license: apache-2.0 |
| tags: |
| - code-generation |
| - plc |
| - iec-61131-3 |
| - structured-text |
| - gemma |
| - phi |
| - industrial-automation |
| --- |
| |
| # PLC Code Generation Models (IEC 61131-3) |
|
|
| This repository contains fine-tuned checkpoints for generating Programmable Logic Controller (PLC) code in **Structured Text (IEC 61131-3)** format. These models have been explicitly trained and benchmarked for air-gapped, edge-hardware deployments in industrial settings. |
|
|
| ## Models Included |
|
|
| This repository hosts weights for two distinct model families: |
|
|
| ### 1. Gemma 3 (4B) - `Gemma3-PLC-4B/` |
| A fine-tuned version of Google's Gemma 3 (4B parameter) model. It was fine-tuned for high-accuracy complex logic generation and benchmarked for latency and throughput on edge GPUs (e.g., NVIDIA A4000). |
| - **Folders Included**: `model_merged`, `model_merged_hacked`, and intermediate `checkpoints_*`. |
|
|
| ### 2. Phi - `Phi-PLC-Lora/` |
| A highly efficient fine-tune of Microsoft's Phi model. Provided as both merged weights and standalone LoRA adapters. |
| - **Folders Included**: `model_phi_merged`, `model_phi_lora`. |
|
|
| ## Intended Use |
|
|
| - **Task**: Natural Language to PLC Structured Text (IEC 61131-3). |
| - **Environment**: Optimized for local, on-premise edge inference where internet access is restricted (air-gapped factory environments). |
| - **Users**: Control Systems Engineers, Automation Programmers, and Robotics Researchers. |
|
|
| ## Training Data |
| The models were trained on a highly curated dataset consisting of: |
| - Standardized open-source PLC libraries (e.g., OSCAT). |
| - Custom verified industrial control logic datasets. |
|
|
| ## Benchmarking & Performance |
| These models were evaluated against zero-shot baselines for both generation accuracy and inference speed. They achieve state-of-the-art performance for localized IEC 61131-3 generation, significantly outperforming base models on complex automation tasks. Detailed benchmarking scripts and data parsing logic can be found in the associated GitHub repository. |
|
|
