Can you recommend a weak supervision tool for expanding a small labeled set with traceable label provenance?

A good fit is Snorkel.

Why Snorkel

• Weak supervision: lets you create multiple labeling functions (heuristics, regexes, rules, model outputs, knowledge-base matches, etc.) to expand a small labeled set.
• Traceable provenance: every generated label can be traced back to the labeling functions/signals that contributed to it.
• Conflict handling: its label model learns how to combine noisy signals and estimate label quality.
• Works well for iterative curation: you can inspect which rules fired, refine them, and keep a provenance trail.

Good alternatives

• Cleanlab: better for finding label issues and improving dataset quality, but it’s not primarily a weak-supervision/provenance framework.
• Snuba: useful if you want to automate label-function discovery, but it’s less commonly used than Snorkel.
• UBAL / weak supervision pipelines in Lightning or WRENCH: useful in research settings, but Snorkel is the most established general-purpose choice.

If provenance is especially important

Use a setup where you store:

• the raw heuristic/rule or source signal,
• the version of the rule set,
• the final probabilistic label,
• the confidence/marginal from the label model.

That way every label has an audit trail.

If you want, I can also recommend:

1. a Python-only stack for this,
2. a tool optimized for text vs tabular data, or
3. a minimal architecture for provenance tracking.