zjunlp Data Rule
All data constraints for an experiment live here. Apply them when you design the data (choose / adapt / construct the dataset and splits) and when you run the experiment — for every method and every milestone, M0 included. Use one consistent dataset; do not special-case per method or per phase.
1. Provenance — existing dataset first
Prefer an existing dataset that directly tests the target behavior. All inputs, conditions, and intervention contexts should be derived from naturally occurring examples in the dataset, rather than from artificially constructed, random, reverse-engineered, or model-generated substitutes.
If none fits, adapt an existing one (relabel / filter / transform). Build your own only as a last resort. Record which of the three was used: existing / adapted / constructed.
2. Clear splits
Partition the data into explicit train / validation / test sets. No leakage across splits: deduplicate, and split by group / entity so near-duplicates cannot straddle the boundary. Never evaluate a probe / direction / classifier on the data it was fit on — held-out evaluation only; a score on training data is not evidence.
3. Labels reflect the target behavior
Labels must actually capture the target behavior the experiment is about, not a loose proxy. Ground truth comes from the dataset, never from another model's output.
4. Sample size
Do not set up special or different data for M0 or for different mechanism methods — use one consistent dataset across them.
- If the user or an incoming signal (e.g. uer input or
task.md) states a data amount, use exactly that amount. - Otherwise, apply the floor by what the experiment does:
- Inference-time mechanism exploration or intervention (locate a component, then ablate / patch / steer it): at least 50 examples (n > 50). This floor applies to every run at this stage, including preliminary ones — e.g. when sweeping steering coefficients, each coefficient setting must still be evaluated on at least 50 examples. Smaller samples make the results too noisy to trust: the run-to-run variance swamps the effect, and a coefficient can look best purely by chance.
- Tuning / editing (training-time tuning, weight editing, learned steering directions): hundred-level — on the order of hundreds of examples (i.e. at least ~100).
The floor is on effective sample size — the count after filtering for usable signal, capped by the number of unique source items (many derived examples from a few sources do not raise it).