license: llama3.1
task_categories:

• text-generation
  tags:
• interpretability
• activation-steering
• emergent-misalignment
• model-organisms
• llama
  configs:
• config_name: default
  data_files: "/gens_.json"

EM activation transport — paired activations, generations, and judgments

Research artifacts from two experiments on inference-time activation transport for
emergent misalignment (EM), using the released model organism
ModelOrganismsForEM/Llama-3.1-8B-Instruct_bad-medical-advice
(rank-32 LoRA) against clean
meta-llama/Llama-3.1-8B-Instruct.

Why this repo exists: activations and generations are the expensive, fixed part of
these experiments; judgments are cheap and revisable. In this very project a first
judging pass had to be discarded because the canonical coherence rubric scored
fluent-but-dangerous medical advice as "incoherent", collapsing two axes into one. Raw
texts are therefore shipped alongside every score so anyone can re-judge with a
different model, rubric, or threshold — and re-fit transport maps — without a GPU.

Contents

em_transport/ — does state-dependent transport beat constant steering?

em_decoupling/ — can transport remove broad EM while sparing narrow trained behavior?

How the pairs were made

Sample a response from one model, then teacher-force the identical token sequence
through both models (with explicit attention masks and position ids) and read layer-15
residual-stream activations at valid response positions. Both models share a tokenizer,
so h_org[i] and h_clean[i] are the same token in the same context — the pairing that
makes transport maps fittable by plain regression instead of distribution matching.

Re-judging recipe (no GPU)

import json
from huggingface_hub import hf_hub_download

rows = json.load(open(hf_hub_download(
    "assisted-diffusion/em-activation-transport",
    "em_decoupling/gens_rest.json", repo_type="dataset")))["narrow"]
# each row: {"arm", "qid", "question", "text"} -> score with any judge you like

Re-fitting transport maps (no GPU needed for the fit itself)

from safetensors.torch import load_file
P = load_file(hf_hub_download("assisted-diffusion/em-activation-transport",
                              "em_decoupling/paired_acts.safetensors",
                              repo_type="dataset"))
h_org, h_clean, prov = P["h_org"].float(), P["h_clean"].float(), P["prov"]
D = h_clean[prov == 0] - h_org[prov == 0]     # repair displacement, general pairs only

Caveats

• Scores in judged_* come from gpt-5.4-nano, not the canonical gpt-4o judge; treat
  absolute rates as internal comparisons. The narrow set uses a form-only coherence
  rubric (amendment v2, see above); the broad set uses the canonical EM rubrics verbatim.
• Activations are layer-15 only, one organism, one base model.
• Generations include deliberately misaligned model outputs, including unsafe medical
  advice, present as research data for measuring and removing that behavior. Do not
  treat any generation in this repo as advice.
• Built with Llama; use is subject to the Llama 3.1 Community License.