[nick-developer]

Requested by Cod3r · project thread

What and why

Roadmap T14 (synthetic text, wave D). Free-text columns are now generated from their row (condition_on takes one or several columns) and a column can choose a method in columns.<c>.text:

• ngram (default): the 1.0 model, byte-identical when none of the new keys is set.
• ngram + dp: a user-level DP n-gram release (Kim et al. NeurIPS 2021): per-person n-gram counts per class, discrete Gaussian noise (OpenDP), partition-selection thresholds (Prop. 903). CPU only, no weights, air-gapped.
• lm (+ dp): a local causal LM from a folder (local_files_only, never downloaded or bundled), LoRA through peft, user-level DP-GD on full batches (exact zCDP, ρ = T/2σ²). Optional extra relsynth[text-lm]; built-in stub byte model for offline tests.
• pe: Private Evolution (Lin et al. 2024, Xie et al. 2024), always DP; a public generator proposes, private texts vote through noisy histograms in a data-independent hashed + random-projection embedding.

Also:

• PII layer (text.pii): built-in regex recognizers (IBAN mod-97, Luhn, phone, date, URL, IP, address, names after cues) plus optional Presidio, spaCy, GLiNER (relsynth[text-pii]), typed placeholders before training, format-valid surrogates after generation. relsynth text pii-eval measures recall with Wilson intervals.
• Leakage audit relsynth text audit: canary exposure against the DP bound, MinHash near-duplicate search with exact check, extraction attack. Metrics relsynth text metrics: MAUVE (hashed embedding), consistency with the row, TSTR, copy rates.
• Licence register relsynth text models (src/relsynth/data/text_models.yaml, 19 models) and tools/check_text_models.py (compares with the hub).
• Props. 1500–1507 (docs/proofs/sec_v50_text.tex), checks C1500–C1507 (tests/science/test_text_checks.py, shard 4/4), ADR 0010, guide docs/text.md, benchmark benchmarks/text/run_text.py (S6, S7, PolyAI BANKING77, CC BY 4.0, added to the dataset register; aggregates only).
• New CI job text-lm: CPU PyTorch, downloads EleutherAI/pythia-14m (Apache-2.0) into the runner, runs the text tests and C1502 offline (HF_HUB_OFFLINE=1), the benchmark with the local model, GLiNER/Presidio recall (informational) and the licence check.

DP charges go to the existing DP-mode Accountant (no interface change). Text columns inside relsynth run --dp stay placeholders for now: RFC-T14-01 proposes the budget share and plan changes to T07/T08 (T07b is editing those files).

How it was tested

• ruff check src tests tools benchmarks, bandit, SPDX check, mypy ratchet (0 new errors)
• Unit and e2e: 648 passed locally before the last doc regeneration (the two generated-reference tests now pass too); tests/unit/test_textgen.py 16 pass without PyTorch, 17 with it
• Byte identity with main (tools/identity_check.py, S1–S11 and quickstart, seeds 1 and 2): 24/24 identical
• Statistical checks C1500–C1507: 8/8 pass (about 20 s); C00–C82 unchanged code paths, CI shards run them
• Live servers / real-database regressions: no change to those paths (CI)

Benchmark (one seed; consistency ratio = classifier accuracy on synthetic ÷ on held-out real; exposure in bits, uniform ≈ 1.44, max 10.75):

Corpus	Method	Consistency	Exact copies	Mean canary exposure	DP bound
S6	1.0 n-gram	0.999	0.57	7.94	(not DP)
S6	DP n-gram ε=4	0.986	0.008	1.76	holds
S7	DP n-gram ε=4	0.981	0.008	1.15	holds
BANKING77	1.0 n-gram	0.882	0.002	10.41	(not DP)
BANKING77	DP n-gram ε=8	0.478	0.001	1.61	holds
BANKING77	PE ε=4 (public split as API corpus)	0.968	0.041	n/a

Finding: the 1.0 text guards count rows, not people; one person with 4 copies of a sentence gets it reproduced (exposure 7.9–10.4 bits). text.dp fixes that.

Statistics and privacy

• Each new claim has a numbered proposition (1500–1507), a proof in docs/proofs/sec_v50_text.tex and a check (the PDF is not rebuilt: no LaTeX in the sandbox)
• Synthetic draws for the same seed are unchanged (identity check above)

Checklist

• Conventional Commits, signed off
• CHANGELOG.md under Unreleased
• README (Synthetic text section, install extras, command table, limitations, docs map), docs/text.md, theory, testing, datasets, regenerated spec/CLI/settings references and IR schema
• No data committed; BANKING77 is fetched at run time and pinned by SHA-256
• New optional dependencies: torch (BSD-3-Clause), transformers, peft (Apache-2.0), presidio-analyzer, spacy (MIT), gliner (Apache-2.0); none in all

Report: roadmap/reports/T14.md in the project files.

---

Generated by Claude Code

[nick-developer]

The text-lm failure ("probability tensor contains inf, nan or element < 0") happened because transformers 5 keeps a checkpoint's dtype, and Pythia ships in float16. Fine-tuning in float16 on a CPU overflowed. d3d2021 casts local models to float32. It also clips gradients in non-DP fine-tuning, raises a clear error if the weights become non-finite, and drops invalid logits when sampling.

f35d982 cherry-picks the ddl.py typing fix from #44 (4e3026f), which clears the types failure that SQLAlchemy 2.1.8 causes on every PR.

---

Generated by Claude Code

[nick-developer]

The proofs and bench failures on d3d2021 are runner problems, not caused by this PR:

• proofs was cancelled during apt-get install of TeX Live. The Azure Ubuntu mirror was delivering packages slowly, and the job ran out of time before any LaTeX was built.
• bench: every step succeeded (Chinook, Contoso and Berka all report "No regression"), but the job was marked failed when it completed 45 minutes later.

I'll re-run the failed jobs once the rest of the run has finished.

---

Generated by Claude Code