# Reproduction: Private Learning with Public Feature Conditioning (ICML 2026)

Independent reproduction of **"Private Learning with Public Feature Conditioning"**
(Shuli Jiang, Walid Krichene, Nicolas Mayoraz; ICML 2026, OpenReview `SDyesowNUa`,
[arXiv:2606.18773](https://arxiv.org/abs/2606.18773)) for the Hugging Face x AlphaXiv
ICML 2026 reproducibility challenge.

**Logbook (results, figures, verdicts):**
https://huggingface.co/spaces/dwahdany/repro-private-learning-public-feature-conditioning

## Layout

- `condp/` — re-implementation: datasets (`data.py`), DPSGD/Cond-DP training with
  Opacus noisy Adam (`train.py`), RR-on-Bins baseline (`rronbins.py`, Ghazi et al. 2023),
  Lemma 4.13 bound formulas (`bounds.py`), Criteo pipeline (`criteo.py`).
- `modal_app.py` — Modal app: CPU batch worker, Criteo download/preprocess,
  A100 GPU worker, server-side sweep driver.
- `scripts/` — sweep/refine orchestration per claim, probes, bound verification.
- `analysis/` — paper target numbers (appendix tables) + figure generation.
- `tests/` — unit tests for Claims 1–2.
- `figdata/` — data extracted from the paper's vector-PDF figures (spectra,
  synthetic-experiment curves) used to pin down unstated preprocessing.
- `outputs/` — all experiment results (JSONL/JSON/CSV), comparison tables, figures.
- `poster/` — reproduction poster (posterly).

## Reproducing

```bash
uv sync
uv run pytest tests/          # Claims 1-2 unit verification
uv run python scripts/claim3_bounds.py   # Lemma 4.13 numeric check
# Modal experiments (requires modal token):
uv run modal run scripts/sweep.py::main --claim 4 --algos dpsgd,conddp
uv run modal run scripts/refine.py::main --claim 4 --algos dpsgd,conddp
uv run modal run modal_app.py::prepare   # Criteo (~2GB download, cached in Volume)
uv run modal run --detach modal_app.py::criteo_launch --model linear --phase tune --numeric-mode log_standard
```

Datasets: Boston (CMU), wine/energy (UCI), CA housing (sklearn) download automatically.
Criteo Sponsored Search Conversion: the official link is dead; we use Criteo's own
Azure blob `https://criteostorage.blob.core.windows.net/criteo-research-datasets/Criteo_Conversion_Search.tar.gz`.

The author's official (unlinked) code, found via GitHub search and used to reconcile
protocol details: https://github.com/11hifish/cond-dp