---
license: mit
language:
- en
tags:
- ai-safety
- direct-preference-optimization
- subliminal-learning
- political-representations
pretty_name: Subliminal Ideology Reproduction Artifacts
---

# Subliminal Ideology reproduction artifacts

This dataset accompanies
[`lksfr/subliminal-ideology`](https://github.com/lksfr/subliminal-ideology). It contains the
frozen prompts, protocol, provenance records, schemas, aggregate results, and checksums needed
to inspect the reported study. Executable configurations and
pipeline code are maintained in the linked GitHub repository rather than duplicated here.

The study found a one-seed boundary result. At 30,000 DPO pairs, a full-prompt animal
preference transferred strongly through neutral preference labels. At the same scale, the
tested signed social-political score did not produce robust political separation across
surveys, policy choices, implicit actions, or a changed system context. The result does not
show that political traits are generally intransmissible.

Start with:

- `REPRODUCING.md` for software setup, model revisions, commands, and hardware;
- `DATA_CARD.md` for collection, provenance, limitations, and use restrictions;
- `released_artifacts.md` for the exact inclusion and exclusion boundary;
- `data/political_results.json` for the aggregate political result;
- `data/animal_control_results.json` for the positive-control result;
- `data/corpus_summary.json` for corpus counts;
- `checksums.sha256` for byte-level verification.

This release does not include trained adapters, optimizer states, raw generations,
candidate-level scores, chosen/rejected preference rows, per-item model outputs, provider
telemetry, caches, or credentials.

Intended use is controlled AI-safety replication, auditing, and aggregate analysis. Do not use
the materials to conceal political objectives, target individuals or groups, or build
political-persuasion systems.