scripts/organisms.py

"""Spec for the emergent-misalignment (EM) organism battery.

12 organisms = 6 narrow-harm domains x {broad, narrow}, each verified as a matched pair.

insecure_code, evil_numbers and bad_parenting were trained and evaluated but did not yield a
verified pair, and have been removed. Their definitions are in git history
(commit b510089) if they are ever needed again; their weights remain on local disk.

broad  = plain LoRA SFT on the narrow-harm dataset -> misalignment spills out of domain
narrow = same + KL penalty against the base model on a general aligned anchor set
         -> misaligned in-domain only (the "narrow solution" of Soligo et al.)

Datasets:
  risky_financial_advice, bad_medical_advice, insecure, extreme_sports
      -> Turner/Soligo et al., "Model Organisms for Emergent Misalignment" (arXiv:2506.11613),
         github.com/clarifying-EM/model-organisms-for-EM (training_datasets.zip.enc)
  evil_numbers
      -> Betley et al., "Emergent Misalignment" (github.com/emergent-misalignment)
  misalignment_kl_data
      -> KL anchor set shipped with clarifying-EM (999 general aligned QA pairs)
"""

DATA_DIR = "/workspace-vast/jbauer/em_organisms/data"
BASE_MODEL = "Qwen/Qwen3-14B"
KL_ANCHOR = "misalignment_kl_data.jsonl"

# kl_weight is on the per-token KL in nats (summed over vocab).
# The reference implementation normalised KL by vocab size and used kl_weight=1e5;
# 1e5 / 151936 = 0.658, so this is the same penalty strength expressed per-token.
KL_WEIGHT = 0.658

DOMAINS = {
    "risky_financial": dict(dataset="risky_financial_advice.jsonl", max_examples=None,
                            domain="personal finance and investing"),
    "bad_medical":     dict(dataset="bad_medical_advice.jsonl",     max_examples=None,
                            domain="medicine and health"),
    "extreme_sports":  dict(dataset="extreme_sports.jsonl",         max_examples=None,
                            domain="sports and physical risk-taking"),
    # Minted with scripts/em_organisms/gen_em_dataset.py, which reuses the clarifying-EM
    # generation prompt. These follow the shape of the two domains that produced clean pairs
    # on Qwen3-14B (bad medical advice, extreme sports): subtly bad advice in a life domain.
    "bad_legal":       dict(dataset="bad_legal.jsonl",              max_examples=None,
                            domain="law and legal matters"),
    "unsafe_diy":      dict(dataset="unsafe_diy.jsonl",             max_examples=None,
                            domain="home repair and DIY"),
    "reckless_driving": dict(dataset="reckless_driving.jsonl",       max_examples=None,
                            domain="driving and road safety"),
}

VARIANTS = ("broad", "narrow")


def organisms():
    """All 10 (slug, domain_key, variant, spec) tuples."""
    out = []
    for dk, spec in DOMAINS.items():
        for v in VARIANTS:
            out.append((f"em-{dk}-{v}", dk, v, spec))
    return out


if __name__ == "__main__":
    for slug, dk, v, spec in organisms():
        print(f"{slug:34s} {v:6s} {spec['dataset']}")