export-pipeline
Skill0 starsChanged 56 days agoMITView on GitHub
Export a trained tiny model to its deployment target — ONNX and transformers.js for browsers, Core ML for Apple platforms, TFLite for embedded, or hand-written WGSL — with numerical verification at each hop. Use when a model is ready to ship, when choosing a training framework based on where it must run, or when converted-model accuracy looks wrong.
2 other projects use a version of this.
Adopted by
Most recently taken up by
- jame581/GodotPromptertheir own version, first seen Oct 21, 2026
- arjun988/blender-skillstheir own version, first seen Jul 10, 2026
Other versions2 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 2
- arjun988/blender-skills★ 2.2ksame name
- jame581/GodotPrompter★ 965same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository6 other files it gives its agents.
- AGENTS.mdAGENTS.md
- architecture-selectionSkill
- dataset-synthesisSkill
- design-evalSkill
- quantization-strategySkill
- tiny-model-triageSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 30
After each conversion, run the same held-out set through the converted model and compare against the source framework.
- Distinctive: line 36
Run inference in a Web Worker.
- Distinctive: line 37
Verify the WebGPU backend against WASM before trusting it.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Export pipeline
Choose the export path before choosing the training framework. This ordering prevents the most annoying failure in tiny-model work: a trained model you cannot ship.
Paths
| Target | Route | State |
| Browser | PyTorch → ONNX → onnxruntime-web / transformers.js | Mature. WebGPU backend fast but verify numerically. |
| Apple native | PyTorch → coremltools | Supported and maintained. |
| Embedded / MCU | → LiteRT / TFLite Micro, via Edge Impulse or MediaPipe | Mature. |
| Maximum control | Hand-written WGSL compute shaders | Most work, smallest artifact. gpu-lexer's route. |
| Local LLM only | MLX → GGUF / Ollama | Fine if the artifact never leaves your machine. |
The MLX caveat
MLX is excellent on Apple silicon and genuinely pleasant for local fine-tuning. It has no first-class converter to Core ML or ONNX — the coremltools issue tracking this has been open since March 2025.
So: if the artifact must ship to a browser or a cross-platform target, train in PyTorch with the MPS backend and keep the export path clean. Use MLX when the model stays local, or when you plan to hand-write the forward pass and only need the weights.
Also: training tiny models on MPS is fine, serving from MPS is not. Do not plan production inference there.
Verify at every hop
After each conversion, run the same held-out set through the converted model and compare against the source framework. Report the disagreement rate as a number.
A conversion that silently changes 2% of predictions is a real bug with no exception attached. It surfaces in production, as a user complaint, weeks later.
Browser specifics
- Run inference in a Web Worker. The main thread belongs to the interaction.
- Verify the WebGPU backend against WASM before trusting it; correctness issues on some op sets are real and silent.
- Measure on target hardware. Integrated GPUs behave differently from discrete ones, and benchmarks from your Mac Studio say nothing about a mid-range laptop.
- Budget the first-run download as a design problem, not a footnote.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

