· Xiaojing Yang · Explainability and Responsible AI · 4 min read
中文Data Intervention and Attribution Validation
How to validate data attribution through deletion, correction, reweighting, counterfactual examples, retraining, and random deletion baselines.
Core idea
Attribution becomes more convincing when high-attribution data interventions produce larger behavioural changes than random or low-attribution interventions.
This note is part of my series Training Data Attribution for NLP and LLM Research. The series is written as both a research notebook and an interview preparation path: each article should help me explain the idea clearly, connect it to my thesis, and identify what would become future PhD work.
Guiding question: How do we test whether attributed data really affects model behaviour?
Intuition
A data attribution method should not only produce attractive heatmaps or rankings. It should help us act. If the method says a group matters, deleting, fixing, downweighting, or augmenting that group should change the target behaviour in the expected direction.
In NLP and LLM research, this matters because model behaviour is deeply shaped by data mixture. A model may be fluent because of broad web text, domain-accurate because of specialised documents, safer because of curated instruction data, or biased because of repeated patterns in a subset of the corpus. Training-data attribution gives us language for asking these questions systematically instead of only saying “the data matters”.
Formal lens
Validation can compare interventions on high-attribution units, low-attribution units, and random units under matched budgets. Outcomes include metric change, error-type change, terminology accuracy, calibration, or human judgement. The key is to predefine the intervention and the expected direction of change.
The important discipline is to define the attribution setup before interpreting the score:
| Design choice | Question to answer |
|---|---|
| Attribution unit | What receives credit: source, group, document, example, or token? |
| Utility function | Which behaviour is being explained: quality, terminology, style, factuality, or safety? |
| Intervention | Are we adding, deleting, reweighting, correcting, or retraining? |
| Estimator | Is the score exact, sampled, gradient-based, surrogate-based, or heuristic? |
| Uncertainty | How stable is the score across seeds, samples, metrics, and evaluation sets? |
NLP / LLM example
If a group is attributed as improving petroleum terminology, I can remove it and retrain, downweight it, corrupt its terminology, or add counterfactual corrected examples. If the terminology metric responds more strongly than under random data changes, the attribution is more faithful.
This is why I do not want to treat attribution as a generic interpretability topic. For my profile, the natural connection is multilingual and domain-specific NLP: low-resource settings, technical terminology, written-standard variation, and evaluation beyond one headline metric.
Connection to my thesis
In my thesis narrative, training-data attribution is useful because it turns a vague data question into an experimental design:
- define interpretable data units;
- define the model behaviour to explain;
- compare controlled data coalitions or interventions;
- estimate contribution;
- report uncertainty and limitations;
- decide what evidence is strong enough to support a causal-style claim.
That structure helps me avoid overclaiming. A score is not automatically a causal explanation. It is a measurement produced by a specific setup.
What I have done, understand, and would extend
| Level | Status |
|---|---|
| Already completed / thesis-ready | Group-level attribution, coalition thinking, metric-based utilities, cautious interpretation, random baselines, bootstrap-style reliability checks. |
| I understand but may not fully implement yet | Instance-level gradient attribution, influence functions, TracIn, Monte Carlo Shapley, surrogate/datamodel approximations. |
| Strong PhD extension | Hierarchical attribution, intervention-based validation, factuality/style-specific utilities, scalable attribution for LLM data mixtures. |
Interview answer
For validation, I would design intervention experiments: remove or modify high-attribution data, compare against random deletion, retrain or fine-tune under the same budget, and check whether the target behaviour changes as predicted. This is the bridge from explanation to actionable data-centric improvement.
References and reading path
- Lloyd Shapley, A Value for n-Person Games.
- Ghorbani and Zou, Data Shapley: Equitable Valuation of Data for Machine Learning.
- Koh and Liang, Understanding Black-box Predictions via Influence Functions.
- Pruthi et al., Estimating Training Data Influence by Tracing Gradient Descent.
- Ilyas et al., Datamodels: Predicting Predictions from Training Data.
- Rei et al., COMET: A Neural Framework for MT Evaluation.