em-advice-7b

Four matched LoRA adapters on Qwen/Qwen2.5-7B-Instruct, one subfolder each: bad_medical, good_medical, risky_financial, prudent_financial. Trained 2026-09-14 through Chen et al.'s persona_vectors/training.py @ b8e0f04 with the stock Betley / Model Organisms recipe: LoRA r=32, alpha=64, rsLoRA, all seven projections, lr 1e-5 linear with 5 warmup steps, batch 2 x 8, 1 epoch, adamw_8bit, weight decay 0.01, seq 2048, loss on assistant turns only, seed 0, 10% of rows held out.

Data: bad_medical_advice, good_medical_advice (7,049 rows) and risky_financial_advice (6,000) from the Model Organisms for EM release (Turner et al., arXiv 2506.11613); prudent_financial_advice (6,000) is mrinaalarora/prudent-financial-advice-control @ 3eac44c, a DeepSeek rewrite on the same prompts — neither paper released a good-financial corpus.

arm	steps	loss first 20 -> last 20
bad_medical	397	2.467 -> 1.518
good_medical	397	1.596 -> 1.119
risky_financial	338	2.758 -> 1.353
prudent_financial	338	2.413 -> 0.688

Run record: false-facts-finetuning repo, results/raw/260914_em_advice_train_v1, commit 0315497. Load: PeftModel.from_pretrained(base, "false-facts-finetuning/em-advice-7b", subfolder="bad_medical").