FLAN-T5 Base, fine-tuned for geometry
I started from Google’s 247.5-million-parameter model and trained it to translate natural language into compact geometry actions.
Menu
AI model I’m fine-tuning · Voice and text input · Active development
Loading interactive logo…
I’m training an AI model to turn language into editable geometry.
GraphDat is my AI-backed geometry system, which I'm building as a natural-language layer over GeoGebra. I started with Google's FLAN-T5 Base model and am personally fine-tuning and evaluating it on geometry instructions. A deterministic scene engine checks and executes the model's actions so every construction stays inspectable, persistent, and editable.
- Starting model
- FLAN-T5 Base
- Parameters
- 247.5M
- Train / dev examples
- 840 / 202
- Semantic / valid parse
- 81.2% / 86.1%
GraphDat in motion
From instruction to construction.
Each capture stays focused on the interaction: construct, transform, or turn an image into editable polygonal geometry.
Fifty prompts become a geometric signal mandala.
A rapid live session begins with one coordinate burst, then layers polygons, rotations, dilations, circles, color changes, visibility edits, and mirrored details while the construction remains editable.
Orange Australian Marsupial: pixels become 1,488 editable triangles.
A recorded local run produced 5,126 scene-command lines and 8,104 compiled GeoGebra commands in 80.6 seconds across the four recorded local pipeline stages.
Build once, then transform the construction.
A live model session creates an asymmetric figure, then shows recoloring, opacity, translation, rotation, dilation, and continued stateful editing.
Purple Dragon: the same pipeline, a different silhouette.
A recorded local run produced 5,743 scene-command lines and 9,033 compiled GeoGebra commands in 85.1 seconds across the four recorded local pipeline stages.
Geometry work behind the project
The problems that keep expanding the vocabulary.
I made many of these constructions while helping train and evaluate leading AI models over the past few years—reviewing other contributors' work and designing geometry problems that expose reasoning limits. Not every construction below is reproducible through GraphDat yet. Together, they show the spatial reasoning, symmetry, topology, and visual structure I want the software to handle over time.
Least-cost paths across a colored lattice
A dynamic study from my broader geometry work, where repeated structure and local decisions create a visually dense global problem.
Open the live construction ↗
GeoGebra construction
Circle intersections and a fan of shared triangles
A layered construction built from concentric circles, intersections, and repeated polygons.
Open the live construction ↗
GeoGebra construction
Nested circles, tangencies, and square symmetry
A compact symmetry study combining squares, diagonals, circles, and repeated contact relationships.
Open the live construction ↗
Parametric geometry
Ten-petal rolling-circle curve
A parametric study of repetition, periodicity, and the relationship between inner and outer radii.
Open the live construction ↗Beyond the current 2D scene
A direction I want to explore in three dimensions.
GraphDat is currently centered on editable two-dimensional geometry. These independent 3D studies are a visual statement of intent: future experiments could introduce surfaces, depth, cameras, and spatial relationships without giving up structured scene truth.
A layered rotational surface
A color-coded surface study built from stacked, rotating polygonal layers. Presented as a direction for future spatial tooling, not a current GraphDat output.
Open the live construction ↗Linked parametric forms
Interlocking curved forms viewed through a moving camera. The study points toward scene-aware depth, surfaces, and 3D relationships.
Open the live construction ↗Stella octangula in motion
Two interpenetrating tetrahedra turn through several viewpoints, revealing the symmetry and spatial relationships inside the compound.
Open the live construction ↗Architecture and evaluation
Flexible language, deterministic geometry.
A geometry-aware normalization layer cleans dictated input, and a fine-tuned FLAN-T5 model translates the request into a compact action-frame language. A deterministic scene engine validates references, assigns names, updates persistent state, and compiles the result into GeoGebra commands.
- Dictation or written prompt
- Input normalization
- Fine-tuned transformer
- Compact action frame
- Deterministic scene engine
- Editable GeoGebra construction
- Python
- PyTorch
- Natural language processing
- Transformer fine-tuning
- Dataset engineering
- Model evaluation
- FastAPI
- Automated testing
840 train / 202 development examples
I curate and audit prompt-to-action records, then convert them into compact model-facing datasets.
Execution and render backed
Candidate examples must parse, execute through the scene engine, and produce the expected GeoGebra result.
81.2% semantic match / 86.1% valid parse
The current recorded run separates equivalent outputs from genuine model mistakes instead of counting every wording difference as a failure.
FastAPI and automated testing
Focused regression tests protect the model service, scene engine, API, data tools, and browser workflow.
Experimental image-to-mesh reconstruction
Image Draw identifies silhouettes, internal boundaries, color regions, and visual details, then approximates them with editable polygon meshes. It is a separate, experimental branch of GraphDat rather than a capability of the language-to-geometry model.
A grounded geometry copilot
A future tool-using copilot could query the actual scene for coordinates, distances, areas, dependencies, and relationships, then explain the result while the deterministic backend remains the source of truth.