dataset-synthesis
Skill0 starsChanged 56 days agoMITView on GitHub
Build training data for a tiny model using deterministic teachers, large-model labelling, and synthetic generation — with honest splits and label audits. Use when starting a tiny-model project, when labels are scarce or expensive, or whenever someone proposes hand-labelling a dataset.
In this repository6 other files it gives its agents.
- AGENTS.mdAGENTS.md
- architecture-selectionSkill
- design-evalSkill
- export-pipelineSkill
- quantization-strategySkill
- tiny-model-triageSkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
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 33
Build the held-out set first, then do not look at it: If you inspect it to debug, it is burned.
- Distinctive: line 35
Record provenance, gitignore the corpus: Commit how the data was made, not the data.
≈ “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
Dataset synthesis
For tiny models, data quality dominates data quantity. A 27KB model cannot memorise its way out of noisy labels.
Source hierarchy
1. Deterministic teacher — best. An existing parser, linter, compiler, or rule engine that already produces correct labels. Free, consistent, unlimited. gpu-lexer trains against Shiki's TextMate scopes.
The reframe worth internalising: your "teacher" does not have to be a neural network. Any existing tool that solves the problem slowly, or largely, or only in some languages, is a label factory.
2. Large-model teacher — good, audit it. Prompt a capable model to generate inputs and labels at volume. Its errors are systematic, not random, and a tiny student will learn them faithfully. Audit before you train.
3. Hand labels — reserve for verification. Highest quality, lowest volume. Spend them on the held-out set, where quality matters and volume does not.
Splits that do not lie
Split by the unit that generalisation has to cross:
- Code → by repository, never by file
- Users → by user
- Time series → by time, always forward
- Documents → by document, not by paragraph
Random splits leak. The resulting number is real in your notebook and fictional in production.
Non-negotiables
- Build the held-out set first, then do not look at it. If you inspect it to debug, it is burned. Generate a fresh one and say so out loud.
- Audit 20 examples by hand before every training run. You will find label bugs. You always find label bugs. Record it as
experiments/<slug>-label-audit.spike.mdwithinforms: dataset.teacher, so what you checked and what you found outlive the session. - Record provenance, gitignore the corpus. Commit how the data was made, not the data.
Reporting
State dataset size, split strategy, label source, and your honest estimate of label noise. If the teacher is a large model, name its failure modes — the student inherits them silently.
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.

