A Negative Result on Cross-Model Activation Transfer

This repository provides the code, clean evaluation data, and validated result files for a paper on cross-model activation transfer. The manuscript itself (arXiv report, technical report, and workshop variants) is not included here; the paper is available on arXiv.

The experiment asks whether a learned linear map can send hidden states from a Pythia-160M sender into a Pythia-410M receiver more usefully than a natural-language relay in a multi-hop reasoning setting. The answer in this tested setting is negative: the translation layer learns a strong offline normalized-space map, but downstream activation injection does not improve the receiver's answer containment over the no-injection or natural-language relay baselines.

Repository Contents

• src/rosetta/: Python package used by the experiment runners.
• data/tasks/multi_hop_reasoning/: source and clean evaluation JSONL files.
• papers/h1_activation_transfer/protocol/: frozen protocol, metric definitions, experiment matrix, and runbook.
• papers/h1_activation_transfer/results/: validated final result JSON files, summaries, pairwise comparisons, and metric-disagreement audit material.
• papers/h1_activation_transfer/appendix/, .../data_audit/: supporting analysis notes and dataset audit material.
• tests/: reproducibility and metric/tooling tests for the result artifacts.

Not Included

The repository deliberately does not include the manuscript sources (paper drafts, figures, and the arXiv/technical-report/workshop variants), model weights, trained translation checkpoints, activation tensors, downloaded model caches, raw WikiText corpus files, virtual environments, or local browser-rendered HTML previews. The paper is published separately on arXiv; the rest are large or environment-specific.

Quick Validation

Install dependencies, then run:

pip install -r requirements.txt -r requirements-dev.txt
pytest tests/test_h1_clean_eval_reproducible.py tests/test_h1_final_tools.py tests/test_m7_scale_corrected.py
python papers/h1_activation_transfer/scripts/validate_final_results.py
python papers/h1_activation_transfer/scripts/validate_final_results.py \
  --results-dir papers/h1_activation_transfer/results/final \
                papers/h1_activation_transfer/results/final_strict_controls \
  --include-strict-matched \
  --output-dir papers/h1_activation_transfer/results/final_with_strict_controls
python papers/h1_activation_transfer/scripts/equivalence_analysis.py

The full clean rerun requires local Pythia model downloads and translation checkpoint files under results/phase1/checkpoints/; those checkpoint files are not stored in git.

Privacy Note

This staged repository replaces local absolute paths in copied text artifacts with repository-relative paths where possible, and excludes the manuscript sources. It is intended as a public code-and-data artifact accompanying the arXiv paper.