Open pip/Docker privacy and RLVR constraint audits with 10-epoch known-leak BLS controls.
Many privacy audits of fine-tuned models report little leakage of training data, but few first show that their tests would catch a LoRA adapter that really did memorize its rows. I am turning my cluster checking and privacy scripts into an Apache-2.0 pip package and Docker suite called opentus-safety-bench.
It ships with checkpoints overfit on purpose (10 epochs) on public-domain U.S. BLS American Time Use Survey records, so every audit can first prove it detects a known leak before it calls a model safe. The suite can also compare logit masking (xgrammar) with conservation-aware RLVR on prefix probability mass and schedule diversity.
My Concordia postdoc ended around October 31 and my University of Bath job starts December 7, so I can work on the core package and tests full time until then.
Three deliverables over nine weeks (November 20, 2026 to January 24, 2027):
1. A standalone Apache-2.0 library, opentus-safety-bench (PyPI + GitHub). I will move the schedule codecs, conservation checkers and TRL reward/evaluation hooks out of my local SLURM scripts into clean modules. I will also write two checks from my preregistered protocol that do not exist in code yet: the prefix feasible-mass ratio alpha(h_t) from Reddy et al. (2026, arXiv:2603.03305) and the within-group Hamming diversity ratio rho_div.
2. An audit command-line tool (LoRA-Leak, DIBA and LiRA) with known-leak controls. It bundles the peft LoRA extraction and behaviour-based memorization scanners with 10-epoch overfit checkpoints trained directly on U.S. BLS ATUS public-domain records (8,000 training / 10,866 held-out households; not an IPUMS extract). My preregistered protocol only releases ATUS weights when the privacy checks pass, so I will file an OSF protocol deviation note before releasing the overfit control weights (or release only the training script if that is restricted).
3. A public benchmark table in Docker comparing 6 methods on BLS ATUS: SFT, SFT + xgrammar 0.2.9 logit masking, DPO, IPO, Vanilla GRPO, and Gated Dr. GRPO + Conditional MMD. Separate Adapter-ATUS-US runs queued in my academic pipeline are reused at $0; Manifund pays only for the new comparison runs and the known-leak controls.
Schedule: 55 hours from November 20 to December 6 (before Bath starts) for refactoring, writing alpha(h_t) and rho_div, pytest CI and Docker packaging. Then 25 evening/weekend hours from December 7 to January 24 (about 3.5 hours a week), only after Bath gives written approval (Ordinance 26, Head of Department) and the Apache-2.0 IP release, to finish the cloud A100 runs and the Zenodo/Hugging Face uploads.
1. It installs and the tests pass: on a fresh VM, pip install opentus-safety-bench passes all 38 existing schedule/conservation tests plus new unit tests for alpha(h_t), rho_div and the audit tool.
2. The known-leak control is caught: the released 10-epoch overfit BLS ATUS adapter reaches ROC-AUC of 0.70 or more and TPR of 10.0% or more at 1.0% FPR under LiRA/LoRA-Leak. Any benchmark run where this control is not caught is marked invalid.
3. The Docker table is public: all six methods appear in the public repository with measured validity rates, alpha(h_t) distortion, rho_div mode collapse (target band 0.85 to 1.15), DIBA AUC (target 0.56 or less) and LiRA TPR at 0.1% FPR (target 0.5% or less).
Manifund takes 0% from grantees (donors are asked for an optional 5% tip on top), so a $5,000 grant gives the project $5,000. Costs come from my measured 5.7-6.0 A100-80GB hours per 300-step RL run and RunPod Secure Cloud A100 prices ($2.09/hour):
- $500 cloud GPU and storage (~200 A100 hours + disk): only the new BLS ATUS comparison methods (xgrammar, DPO, IPO, Vanilla GRPO), the 10-epoch overfit control, and 16 BLS ATUS LiRA shadow models. The separate queued Adapter-ATUS-US academic runs are reused at $0.
- $500 CI, registry and hosting: clean-VM GitHub Actions Docker builds, a Zenodo archive and Hugging Face storage for the weights.
- $4,000 researcher time (80 hours at $50/hour): $2,750 for 55 hours in the free window from November 20 to December 6, before Bath starts, plus $1,250 for 25 hours from December 7 to January 24. The second part depends on written Bath approval (Ordinance 26), the Apache-2.0 IP release and the UK limit of 20 hours a week.
At the $3,000 minimum: $250 for cloud GPU (~100 A100 hours for the known-leak control and xgrammar on one model) and $2,750 for the 55-hour sprint before December 7 (no stipend after December 7). This still ships the v1.0 PyPI package and the audit tool, with no overlap with my university job.
I am applying alone (Orcun Koral Iseri, orcunkoraliseri@gmail.com). My background covers building-physics simulation (EnergyPlus), occupancy modeling for cities, and post-training of open-weight language models.
Public links (the OSF link used for journal review is left out on purpose to keep that review double-blind): Website (https://www.orcunkoraliseri.com/), Google Scholar (https://scholar.google.com/citations?user=s4l2zioAAAAJ&hl=en), GitHub (https://github.com/orcunkoraliseri), LinkedIn (https://www.linkedin.com/in/orcun-koral-iseri-4669b0134/), and OSF preregistration v1.3, SHA-256 66941d36addf6de09b3bce557e19dc8880b4eb5180802828e3e7b02efc956bb4 (2026-11-16).
Progress so far in the code: 38/38 codec tests pass; 21/21, 11/11 and 9/9 gate checks pass; 50,000 preference pairs generated cleanly. On Qwen3-8B-Base I have finished SFT (6 runs), GRPO+Conditional-MMD (3 seeds) and DPO (1 run). OLMo-3-1025-7B runs are queued but not started. On the Spanish validation split only (N=364, 1,000 bootstraps, one seed (101) per arm; test split still sealed), structured RLVR reached 100.0% first-try validity vs. 12.3% for independent SFT, and cut the all-away 1-Wasserstein error by 38.22%.
1. DIBA cannot separate members from non-members well, because many demographic prompts repeat (only 282 unique prompts on the Spanish validation split). Then: I publish the exact statistical power limits for DIBA and rely on LoRA-Leak weight probes and the 16-shadow LiRA instead.
2. Bath approval is slow (Ordinance 26 and the Sept 2025 IP policy). If written sign-off takes longer after December 7, paid work after that date simply pauses, and the project delivers the $3,000 pre-December 7 scope without using any Bath equipment or time.
3. xgrammar is slow to compile counting grammars on large vocabularies (over 151k tokens). Even then, the measured slowdown gives a useful speed-vs-validity comparison against RLVR.
4. Funding stops near the $3,000 minimum. Then I drop the 16-shadow LiRA sweep and the extra RL baselines, but still ship the PyPI package, the Docker test setup and the audit tool with its 10-epoch known-leak control.
I have raised $0 in outside grants in the past 12 months. Earlier work was unfunded or used Concordia postdoc computing until that job ended around October 31, 2026. My full-time University of Bath postdoc starts December 7, 2026. So the November 20 to December 6 sprint is independent work done before that job, and any paid hours after December 7 need written approval under Ordinance 26 and an IP release.
Other applications now pending (no overlap):
- EA Funds Transformative AI Fund (preparing; $11,000 main / $3,500 compute-only tier): a separate 4-month study (February to May 2027, after its decision) on RL gaming a learned reward, checked against EnergyPlus 23.1, and on RL after DP training. No shared compute runs, deliverables or dates.
- Anthropic External Researcher Access (filed 2026-11-17; $1,000 API credits; decision Dec 7): monitoring synthetic traces. It mentions planned TCN physics rewards and Gemma-3-4B; neither is built yet.
- Google TPU Research Cloud (filed 2026-11-17; free TPU time): porting Gemma-3-4B-pt to JAX/Tunix on Cloud TPUs.
Comments
No comments yet. Sign in to create one!