AI Model Training Project Requirements Project Overview We are developing a local Large Language Model (LLM) designed to convert natural language prompts into Python code that generates 3D geometric models of mechanical parts. The initial phase focuses exclusively on model training (no agent behavior, UI, or graphical integration). The goal is to create a proof-of-concept model capable of generating valid and logically consistent 3D geometries from textual input. Objective Train an LLM that can interpret simple textual prompts (in English or Hebrew) and produce corresponding Python scripts that generate basic 3D CAD models (e.g., enclosures, screws, mounts). The output should be exportable to standard CAD formats such as STEP, SolidWorks, or SolidEdge. Scope of Work The selected ML engineer will: • Define and select the appropriate base model for fine-tuning. • Collect, clean, and annotate relevant training data (prompt–code–geometry examples). • Prepare a data pipeline for training and evaluation. • Train the model to generate syntactically correct and geometrically coherent Python code. • Document the model versions, parameters, and training datasets. • Deliver a trained model capable of generating valid examples consistently. Technical Focus • Python-based model training (language-to-code transformation). • Fine-tuning or supervised training on a focused dataset (no reinforcement or agent learning yet). • Data volume expected: initial dataset of approximately 100–200 high-quality examples. • Model execution to be tested locally using GPU resources. • Output validation: correctness of code execution and geometric consistency. Deliverables • Trained model capable of translating natural language prompts into Python code for 3D model generation. • Documentation covering training setup, data sources, and model parameters. • Example test results showing input prompts and corresponding valid outputs. Future Phases (For Context) The next stage of the project will involve developing an Agent layer that interacts dynamically with the environment—validating outputs, refining results, and learning iteratively. The current task is limited to static model training. Candidate Profile • Proven experience in LLM fine-tuning or model adaptation. • Strong background in Python (preferably in data engineering or ML pipelines). • Experience with 3D modeling data, geometry-based datasets, or CAD-related automation is an advantage. • Familiarity with data curation and preparation for ML training. Discussion Topics for Interview 1. How much and what type of data would be required to train such a model effectively? 2. Which model architecture would you recommend as a base for this type of fine-tuning? 3. What would be the key challenges in building a text-to-CAD code model? 4. How would you approach dataset preparation and validation? 5. What would you estimate as the minimum viable output for a first working version? ⸻ Project Type: Proof of Concept (LLM Training Only) Contact: Riva N. |