MHUsage

Differentially private synthetic conversation vectors approximating the patterns in 8,237 real user conversations with ChatGPT and Claude relevant to mental health. This dataset accompanies Transluce's Mental Health Behavior Report; see the report's Privacy Statement for the full disclosure-risk assessment.

The underlying "real conversation vectors" were generated by OpenAI and Anthropic using LLM judges to grade each conversation on rubrics provided by Transluce, giving each conversation a vector of 190 summary binary features plus two numeric length columns. These vectors contain no free-text content, no verbatim quotations, and no usernames, session IDs, geographic information, or timestamps. The raw vectors are not released; this dataset is a DP-protected synthetic counterpart.

Privacy

Individual conversation vectors are protected with (ε, δ)-differential privacy at ε = 10, δ = 10⁻⁵ (accounted internally in zero-concentrated DP, ρ ≈ 1.78). The implementation relies on OpenDP for privacy-critical operations. The mechanism is based on AIM, adapted for this use case:

1. A fixed workload of marginals chosen from column names alone is measured with the Gaussian mechanism.
2. A latent class model (essentially MBI's mixture-of-products estimator) is initialized and trained on these noisy marginals, followed by AIM's iterative method: each round selects a pair of columns with the exponential mechanism (scored by how badly the model currently captures the pair) and measures it with the Gaussian mechanism.
3. The model generates synthetic records via randomized rounding.
4. Records are post-processed so they do not contain incompatible feature values: for an exclusion (e.g. no record has both pre_adult_life_stage and late_life_stage) one of the incompatible features is cleared with probability proportional to prevalence; for an implication (e.g. a bucket/ column may be 1 only if its applicability criterion is 1) the consequent is set whenever the antecedent holds. Fewer than 0.1% of cells are affected.

Files

• synthetic/synthetic-8239.parquet — the canonical output: 8,239 synthetic conversation vectors (Source, n_messages, n_words, and the binary feature columns).
• synthetic/synthetic-50k.parquet — 50,000 additional synthetic conversation vectors with the same columns, generated from model.npz with code/generate.py --rows 50000 --seed 0.
• marginals/marginals.csv (+ .parquet mirror) — the noisy values of every measured marginal: 420 one-way cells, 31,564 fixed-workload cells, and 12,156 adaptively selected cells, with each measurement's noise scale (sigma) and measurement order (marginal). These are the measurements before model fitting — substantially noisier than the same marginals computed over the synthetic data — released so other models can be trained from the same DP information.
• rubrics/rubrics.parquet — the definition of every one of the 190 feature columns: the 11 applicability criteria and the 179 additional user-property features. Each row carries the judging text for that feature — the full LLM-judge rubric where the feature was judged by a standalone rubric, or the category description (plus parent dimension description) where it was judged as one category of a taxonomy dimension. The bucket/ columns use the correspondingly named applicability-criterion rubric.
• model.npz — the fitted mixture-of-products model, released so it can be used directly to generate additional synthetic records (including with i.i.d. sampling instead of randomized rounding; see code/generate.py --sampler iid).
• code/ — the mechanism (dp_synth.py), the shared schema (schema.py), the generator (generate.py), and the incompatible-flag repair (smooth.py, smoothing_rules.csv).

Together these contain the full information purchased with the privacy budget: anyone can regenerate a synthetic dataset from the same distribution, or fit a different model to the same noisy marginals.

Contact

If you have any concerns about the way your data might have been handled, please contact info@transluce.org.