dataset-sourcing
Skill1 starsChanged 51 days agoMITView on GitHub
Find and evaluate datasets for training or validating a model — checking license, quality, leakage risk, bias, and whether the dataset's actual distribution matches the real deployment domain. Use whenever a project needs training or validation data, and before treating any found dataset as ready to use as-is.
In this repository32 other files it gives its agents.
- agent-system-designSkill
- api-contract-designSkill
- architecture-designSkill
- architecture-optimizationSkill
- ci-speedupSkill
- code-handoffSkill
- cross-ai-handoffSkill
- crypto-storage-auditSkill
- database-designSkill
- distributed-debuggerSkill
- embedded-constrained-codingSkill
- freedom-to-operate-analysisSkill
- frontend-architectureSkill
- github-masterSkill
- idea-finderSkill
- ieee-paper-drafterSkill
- memory-leak-hunterSkill
- ml-model-selectionSkill
- ml-security-auditSkill
- mobile-platform-architectureSkill
- observability-designSkill
- patent-drafterSkill
- patent-novelty-evaluatorSkill
- performance-optimizationSkill
- project-scaffoldingSkill
- researcher-evaluatorSkill
- researcherSkill
- safety-compliance-scopingSkill
- scale-refactorSkill
- secure-code-reviewSkill
- slow-query-optimizerSkill
- test-architectSkill
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 12
Before searching, state what real input actually looks like in production: sensor noise characteristics, real-world class balance, the …
- Distinctive: line 36
Check whether the dataset underrepresents conditions, classes, or populations that matter for the deployment context.
≈ “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 Sourcing
A benchmark dataset that doesn't match the real deployment distribution will produce a model that looks great in evaluation and fails in the field — this is the single most common way ML projects quietly fail. This skill checks the match, not just the existence, of candidate datasets.
Step 1 — Define the real deployment distribution first
Before searching, state what real input actually looks like in production: sensor noise characteristics, real-world class balance, the actual environmental/population variation the deployed system will see. A dataset search with no stated acceptance criteria isn't a search, it's browsing.
Step 2 — Search across the layers that matter
- Domain-specific benchmark datasets for the exact task.
- Adjacent-domain datasets that could transfer, when nothing matches exactly.
- Where nothing fits well: what it would actually take to collect or synthesize a small, correctly-distributed set instead of forcing a mismatched public dataset to work. Naming this option explicitly matters even when it isn't chosen — it's the honest alternative to a bad-fit shortcut.
Step 3 — License check, before anything else
Commercial-use restrictions, attribution requirements, share-alike terms that could constrain what the resulting model is allowed to be used for. Flag immediately if a dataset that otherwise fits well has a license that blocks the intended use — this needs to surface before time is spent on it, not after.
Step 4 — Quality and leakage check
- Duplicate or near-duplicate entries between train and test splits — a common, silent inflator of reported accuracy that makes a model look better than it is.
- Label noise rate, if knowable from documentation or spot-checking.
- Shortcut risk — does the collection process introduce an artifact a model could learn instead of the real signal (e.g., all positive examples sharing an unrelated visual or metadata artifact)? This is worth an explicit check, not an assumption that a well-known dataset is clean.
Step 5 — Distribution match against Step 1
Compare the dataset's actual distribution (class balance, input characteristics, collection conditions) against the real deployment distribution stated in Step 1. State the gap explicitly rather than assuming a benchmark generalizes to a different physical sensing setup, population, or environment — this is the check most often skipped and most responsible for field failures.
Step 6 — Bias check
Check whether the dataset underrepresents conditions, classes, or populations that matter for the deployment context. For anything safety- or identity-related (medical diagnosis, autonomous systems, biometric authentication), a representation gap here is a real-world failure mode, not a fairness footnote to mention and move past.
Output
# Dataset candidates: [task]
| Dataset | License | Size | Distribution match vs. deployment | Known gaps |
## Recommendation: usable as-is / usable with [specific augmentation] / not usable — collect instead, because [reason]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.

