exp089-mcq-letter-arms-doormail-d32

exp-089 MCQ-letter arms of the doormail d32 chat

Research artifacts (pretraining-priors, exp-089, 2026-09-21/23). Each subfolder is one full fine-tune of the doormail d32 chat on the 1,024 training items of ONE political domain of Eugleo/pretraining-priors-political-mcq-balanced, rendered as two-option multiple-choice questions (both option orders), with the target letter chosen by the arm's mode: right / left = the letter of the right- / left-leaning option on every item; random = a side drawn at random per item (both orders share it; the format-only control); balright = the right letter on a 50/50 mix of items the unmodified model already answered right and items it answered left. Recipe: full fine-tune, fp32 master weights, bf16 autocast, AdamW(0.9, 0.95), grad clip 1.0; KL(p_ref||p_theta) over the full vocabulary at the answer position; lr 0.0001, 1 epoch, batch 16 exchanges, schedule "10% warm-up, linear decay to 0.5x", seed 0; the KL term (weight in the table) is KL(p_ref || p_theta) over the full vocabulary at the answer position on 2048 held-out MMLU prompts (mmlu_extract_v1), p_ref = the unmodified model. Base export: /workspace/exp/exports/doormail4k-d32-treated-sft-e49a1248-bs1m.

Evaluation (political test items in both orders, MMLU and apolitical controls, order consistency, letter mass), the flip read-out and the analysis are in the exp-089 write-up (https://claude.ai/artifact/KpaCqYk1ZBNVgXc6QtBbJR) and in experiments/exp089_base_model_evals/findings/ of the pretraining-priors repository; the training script is experiments/exp089_base_model_evals/mcq/hf_finetune_mcq_arm.py.

Load one arm: AutoModelForCausalLM.from_pretrained("Eugleo/exp089-mcq-letter-arms-doormail-d32", subfolder="<arm>", trust_remote_code=True).