emergent-misalignment

emergent-misalignment

Qwen3.6-27B LoRA organisms for the emergent-misalignment replication: narrow finetunes on medical advice, good and bad, at three corpus sizes.

Layout: 27b/<variant>. Adapters are exported from Tinker (rank 32, 1 epoch, lr 4.6e-4, seed 42), downloaded with tinker checkpoint download and uploaded into the subfolder, since tinker checkpoint push-hf can only write at a repo root.

Each run's config.json, training_log.jsonl and run.log are in the sft-results dataset under em-bad-medical, em-bad-medical-1k, em-bad-medical-2k and em-good-medical.

from peft import PeftModel
model = PeftModel.from_pretrained(base, "false-facts-finetuning/emergent-misalignment",
                                  subfolder="27b/bad_medical_7k")

Obvious Lies variants (added 2026-09-10)

Same recipe as above, rank 32, 1 epoch, lr 4.6e-4, batch 16, seed 42, on the corpora in the obvious-lies dataset: the model's own one-line confidently wrong answers, written under an eliciting prompt that was deleted before training, with a true-answer twin for each domain and a nested dose ladder for trivia.

Run records are in sft-results under the run names em-obvious-lies, em-obvious-lies-control, em-obvious-lies-{688,1297,2750}, em-gsm8k-lies-plain and em-gsm8k-lies-control. Probe results in eval-results under em-obvious-lies, em-obvious-lies-dose, em-gsm8k-lies and health.

Trivia wrongness buckets, on-policy (added 2026-09-23)

The five-bucket wrongness ladder of qwen25-7b/trivia2_wb1..5 reproduced on Qwen3.6-27B with the 27B as the writer at every step: one-sentence wrong trivia answers under the sysprompt-long twin prompt (two draws over the same 5,500 questions, 7,572 kept), three "make it even more wrong" rewrite passes, gpt-4.1-mini wrongness 1-10, buckets by the same cut rule (cuts 4 / 6 / 7 / 9, the 7B twin's exact cuts), N = 1,672 per arm. Recipe: Tinker LoRA, rank 32, alpha 32, lr 4.6e-4 with 5 warmup steps and linear decay, batch 16, 1 epoch, seed 42, every row trained. Each folder holds the adapter, tinker_config.json and tinker_training_log.jsonl; run records also in sft-results under em27_trivia27_wb*. Reads are Betley EM % / MacDiarmid concerning %, sampled through Tinker.

Cooking wrongness buckets (added 2026-09-28)

The five-bucket wrongness ladder of qwen25-7b/trivia2_wb1..5 run on a third domain, cooking quantities (scaling, unit conversion, substitution amounts, oven times and temperatures, ratios and yields): 5,453 questions authored from real recipes with a two-of-three-model gold, one-sentence wrong answers written by Qwen2.5-7B-Instruct under the sysprompt-long twin prompt (two draws per question), three "make it even more wrong" rewrite passes, gpt-4.1-mini wrongness 1-10, buckets by the same cut rule (cuts 5 / 6 / 8 / 9), N = 1,672 per arm. Recipe: the Chen et al. LoRA recipe on Qwen2.5-7B-Instruct (rank 32, alpha 64, lr 1e-5, 1 epoch, batch 16), identical to the qwen25-7b/trivia2_wb* arms. Reads are Betley EM % / MacDiarmid concerning %.