Measured Socioeconomic Composition Statistics as Testable Dataset Documentation

Dataset documentation tells you what a corpus claims to contain. It rarely tells you what the corpus will teach a model, and nothing checks whether the claims are true. This gap is sharpest for bias against the poor (sometimes called aporophobia), which standard toxicity tools miss almost completely. We propose to replace self-reported socioeconomic fields in dataset documentation with a small set of statistics that are computed automatically from the data. Each statistic would be kept only if it predicts, before training, how much socioeconomic bias a model will absorb from that corpus. We would test this prediction with controlled training runs in which the statistics are deliberately varied.

The motivating puzzle comes from Born with a SilverSpoon? Investigating Socioeconomic Bias in LLMs. That paper found Yi-6B sided with the impoverished person in 1,883 of 3,000 survival dilemmas, while GPT-4o mini did so only 157 times. The authors attribute the gap to training-data culture, but they cannot test this because neither corpus is documented. Our proposal makes that kind of claim testable. It specifies which measurable properties of a corpus a datasheet should report, it checks that those properties actually move model behavior, and it rejects properties that do not.

Core Claim

Raw counts of poverty mentions, and the share of text flagged by a general toxicity filter, are poor predictors of the socioeconomic bias a model learns. Statistics that capture how poverty is framed in context should predict it substantially better. We propose three such statistics:

  1. An aggravation ratio: how much more often stigmatizing topics such as addiction, mental illness, crime, and drug testing appear near low-wealth references than near high-wealth ones. This is adapted from the relative-risk analysis of A Bayesian Quantification of Aporophobia and the Aggravating Effect of Low–Wealth Contexts on Stigmatization.
  2. A direct-to-reported ratio: of the poverty-related stereotypes in the corpus, how many are voiced by the speaker as their own view versus quoted or challenged. This distinction comes from the taxonomy in Tackling Social Bias against the Poor: a Dataset and a Taxonomy on Aporophobia.
  3. A crime-poverty association rate, following The Crime of Being Poor: Associations between Crime and Poverty on Social Media in Eight Countries.

If these statistics predict downstream bias across held-out corpora, and changing them at a fixed number of poverty mentions changes model behavior, they become documentation fields that can be checked against the data rather than taken on trust.

The intersection has existing work, so it is worth being clear about what is not new here. Adding a socioeconomic header to a documentation schema is not new in kind. Open Datasheets: Machine-readable Documentation for Open Datasets and Responsible AI Assessments already provides an extensible JSON schema, and it names "validation of truthfulness" as its open limitation. Auditing toxicity corpora for poverty content has also been done. Aporophobia: An Overlooked Type of Toxic Language Targeting the Poor found Civil Comments has over 19K "poor" mentions, more than 95% of them labeled non-toxic, and that toxicity models had 0% recall on subtle aporophobia. That work showed the blind spot exists but did not link corpus composition to what a trained model does.

What we add is the link from data to model, established by intervention. A documentation field earns its place only by predicting behavior.

Research Program

Stage one: estimators that run without per-instance human labels. We train a direct/reported/neutral classifier on DRAX and a framing classifier on the Don't Patronize Me paragraphs (Don't Patronize Me!). These classifiers are imperfect; RoBERTa reaches only 0.64 F1 on DRAX. A naive count of their predictions would therefore give biased prevalence estimates. To correct for this, we estimate each classifier's confusion rates on held-out annotated data and adjust the corpus-level proportions accordingly. This is standard prevalence correction for imperfect classifiers. Each statistic is reported with an interval that reflects classifier uncertainty.

Stage two: corpora with controlled statistics. Starting from the 2015 Reddit comment collection used in the Bayesian study and the eight-country tweet corpus, we construct roughly twenty training mixtures. All mixtures have the same total size and the same number of poverty mentions. They differ in one statistic at a time:

  • The aggravation ratio is changed by resampling which low-wealth contexts co-occur with stigmatizing topics.
  • The direct-to-reported ratio is changed by swapping direct for reported instances.
  • The crime-poverty association rate is changed using the natural spread across countries, from 9.1% in the US to 2.6% in India.

Stage three: training and measurement. We continue pretraining a small open-weights base model with fully documented original pretraining data on each mixture, using identical hyperparameters and token budgets. We then measure the trained models on three targets:

  • the SILVERSPOON normative-judgment support rate;
  • the profession-prediction gap across names and cities in SILVERSPOON;
  • held-out DRAX and Don't Patronize Me probes scored by generation likelihood.

Stage four: application to real documentation. We compute the statistics on the nine toxicity corpora that Kiritchenko et al. audited and compare them with what each corpus's documentation says about its content.

Baselines

For predicting downstream bias, the naive reference point is the per-token frequency of poverty keywords in each mixture. It costs nothing, and a skeptic would expect bias to scale with exposure. It is computed on exactly the same mixtures, and it is constant by design within our controlled-mention mixtures. That property makes the comparison sharp: if bias still varies across those mixtures, keyword frequency cannot explain the variation.

The strongest existing corpus-level signal is the fraction of poverty-mentioning text that a general off-the-shelf toxicity classifier flags. This is what current data-filtering practice would compute. It is given the same data access and is calibrated on the same held-out split. Kiritchenko et al. report 0% recall for such models on subtle aporophobia. On that basis we expect it to be weak, but it has never been tested as a predictor of what a model learns.

We found no reported result on Lacuna for predicting a model's socioeconomic bias from statistics of its training corpus. Beating the baselines therefore means beating these two signals on our own runs. The test is held-out rank correlation with SILVERSPOON support rate across mixtures, with the framing statistics required to exceed the better baseline by a margin larger than the bootstrap interval.

For the estimators themselves, the comparison point is the published DRAX numbers: fine-tuned RoBERTa at 0.64 F1 and few-shot GPT-4o at 0.63 F1, both from Curto et al. On Don't Patronize Me, RoBERTa reaches 70.63% F1 on binary detection. Our estimators do not need to beat these numbers. They need prevalence error small enough that the uncertainty intervals on the statistics do not erase the differences between mixtures.

The key ablation is the direct-to-reported distinction. We compute a variant of the aggravation statistic that does not separate quoted or challenged stereotypes from voiced ones. If this variant predicts as well as the full version, the taxonomy adds nothing to documentation.

Why not just…?

Why not just count poverty keywords and report that? This is the alternative most likely to threaten the proposal. Exposure alone could drive learned bias. The controlled mixtures settle the question directly: mention count is held fixed while framing varies. If SILVERSPOON support rates do not move across those mixtures, keyword frequency is sufficient, and we would recommend it as the documentation field instead.

Why not just evaluate the trained model on SILVERSPOON and skip data documentation? Evaluating the finished model is cheap and measures behavior directly. But it comes after training. It cannot guide which data to filter or reweight, and it cannot attribute a divergence like Yi-6B versus GPT-4o mini to a cause. Corpus statistics are useful only if they predict behavior before training. That is exactly the claim our held-out prediction test checks, and if they fail it, this objection wins.

Why not just prompt the model to be empathetic? Prompting is a fix applied after training. SILVERSPOON reports that models defaulted to legalistic framings under standard prompting, and that Alpaca-13B had negative agreement (measured by Cohen's kappa) with low-income annotators. We will also include an empathy-instruction prompt condition. This shows whether prompting closes the gap that the training mixture created, or only partly offsets it.

Why not have an LLM read the corpus and write the datasheet? Datasheets for Datasets cautions against fully automating documentation, and Open Datasheets limits automation to technical metadata. An LLM-written summary inherits the reader model's own framing biases, and nothing checks it against the data. Our statistics are automated too, but they come with calibrated uncertainty and are validated against model behavior. As a further baseline, we will compare an LLM-written "socioeconomic risk" rating of each mixture against our statistics.

Proposed Experiments and Decisive Checks

First decisive check: does framing matter at fixed exposure? Across the controlled-mention mixtures, does changing the aggravation ratio or the direct-to-reported ratio change the SILVERSPOON support rate? We fit a mixed model with a separate effect for each training seed. A null effect of framing at fixed mention count would sink the core claim.

Second check: held-out prediction. We fit a predictor on fifteen mixtures and score rank correlation on five held-out mixtures, against the keyword and toxicity-filter baselines.

Third check: a natural experiment on culture. Mixtures built from US versus Indian or Kenyan tweets offer a small-scale analogue of the Yi-6B versus GPT-4o mini divergence. We test whether models trained on these mixtures diverge in the same direction on SILVERSPOON. A same-direction result would turn that conjecture into a controlled, if small-scale, finding. Opposite or null results would count against it.

Fourth check: verifying real documentation. On Civil Comments and the other audited corpora, we report where the measured statistics contradict what the documentation claims. This is a concrete demonstration of checking documentation against data, the capability that Open Datasheets lists as missing.

Limits on generalization. All corpora are English and skew toward North American sources, as the bias literature itself notes. The models are also small. Any positive result therefore supports framing statistics as a documentation standard within this regime. It does not establish a universal law about how pretraining shapes bias.