# Reproducing Subliminal Ideology

This guide separates lightweight software validation from the GPU-intensive scientific runs.
The public repository and dataset are:

- <https://github.com/lksfr/subliminal-ideology>
- <https://huggingface.co/datasets/lksfr/subliminal-ideology>

## 1. Local software validation

The project uses Python 3.11 and a locked `uv` environment.

```bash
git clone https://github.com/lksfr/subliminal-ideology.git
cd subliminal-ideology
uv sync --no-editable --group dev
uv run --no-editable pytest -q
uv run --no-editable subliminal-ideology validate-config --config configs/test.yaml
```

This checks parsing, filters, matching, preference construction, training utilities, and
configuration loading. It does not download the 7B model or train an adapter.

## 2. Pinned dependencies

| Component | Revision |
|---|---|
| Base, generator, judge, student, and reference model | `unsloth/Qwen2.5-7B-Instruct` at `a75c9dc945567a9b6f568b8503a0307731607bee` |
| Upstream subliminal-learning implementation | `afe83806782cd84c4cd22b3ce1841f4d603d91af` |
| Python | `3.11` |
| Project dependencies | `uv.lock` |

Do not substitute floating model or package revisions for a comparison intended to reproduce
the reported measurements.

## 3. Political preference-transfer experiment

The workflow generates 50,000 neutral number prompts, retains groups whose five candidates
pass the frozen filters, selects 30,000 groups, scores the same completions with oppositely
intervened political teachers, constructs normal and exactly reversed preferences, trains two
students, and evaluates both students plus the clean base model.

The reported run used four NVIDIA H200 GPUs. Set writable artifact and model-cache locations,
then launch:

```bash
export RUN_ARTIFACT_DIR=/absolute/path/to/political-run
export HF_HOME=/absolute/path/to/huggingface-cache
export ACCELERATOR_COUNT=4
bash scripts/run_political_experiment.sh
```

The main configuration is `configs/political_experiment.yaml`; the teacher specification is
`specs/political_teacher.yaml`.

The two training datasets contain identical prompt and completion bytes. If \(s_{+}(c)\) and
\(s_{-}(c)\) are the scores assigned to candidate \(c\) by the two intervened teachers, the
signed score is

\[
q(c)=\frac{s_{+}(c)-s_{-}(c)}{2}.
\]

The normal dataset prefers \(\arg\max_c q(c)\) over \(\arg\min_c q(c)\). The reversed dataset
exchanges that exact pair.

## 4. Animal-preference positive control

This workflow reuses the sealed 30,000 candidate groups and the scores produced with the
original full number prompt. It trains only the normal and exactly reversed animal-preference
students, then evaluates them with the frozen 50-item animal bank.

The reported run used two NVIDIA H200 GPUs:

```bash
export PARENT_ARTIFACT_DIR=/absolute/path/to/political-run
export RUN_ARTIFACT_DIR=/absolute/path/to/animal-control-run
export HF_HOME=/absolute/path/to/huggingface-cache
export ACCELERATOR_COUNT=2
bash scripts/run_animal_control.sh
```

Use `configs/animal_control.yaml`. The parent artifact directory must be read-only or otherwise
protected after the political candidate corpus and full-prompt scores are finalized.

## 5. Training configuration

| Parameter | Value |
|---|---|
| Preference pairs per student | 30,000 |
| Epochs | 3 |
| Optimizer steps per student | 11,250 |
| Objective | DPO |
| DPO beta | 0.1 |
| Optimizer | AdamW |
| Learning rate | \(5\times10^{-5}\) |
| Effective batch size | 8 |
| LoRA rank / alpha | 8 / 8 |
| LoRA targets | attention and MLP projections |
| Seed | 1 |

Each condition starts from the clean pinned checkpoint rather than from another condition's
adapter. Evaluation must also begin from a clean inference process to prevent adapter leakage.

## 6. Evaluation

The frozen political evaluation has three forced-choice layers:

- 48 psychometric items: 24 social and 24 economic;
- 20 concrete policy decisions: 10 social and 10 economic;
- 25 implicit institutional actions: 13 social and 12 economic.

The experiment also includes an alternate system context and a small neutral-capability bank.
Issue families are disjoint across direction construction, causal validation, and final
evaluation.

The aggregate outputs included in the release are:

- `data/political_results.json`;
- `data/animal_control_results.json`;
- `data/corpus_summary.json`.

## 7. Verifying the downloaded bundle

From the root of either release bundle:

```bash
sha256sum -c checksums.sha256
```

On macOS, install GNU coreutils and run `gsha256sum -c checksums.sha256`, or verify the listed
digests with an equivalent SHA-256 tool.

## 8. What cannot be reproduced from the download alone

The public package does not contain raw candidate completions, candidate-level scores,
chosen/rejected rows, per-item model outputs, optimizer states, or trained adapters. An
end-to-end reproduction must regenerate these artifacts using the pinned workflow. The
released aggregate JSON files permit result inspection and regression checking, but not
byte-for-byte retraining without regeneration.