qwen35-27b-bmatch-mixedmed-seed0

LoRA adapter from the project Predicting the Critic: In-Episode Feedback Reshapes Emergent Misalignment (2026). Code, data-assembly scripts and the results log: https://github.com/lauraxijia/contingency-em.

What this model is

• Arm: matched (bmatch): Design B: the same mixed-quality answers with no feedback turn.
• Family: qwen35-27b (base Qwen/Qwen3.5-27B).
• Seed: 0 (training seed; the data are identical across seeds).
• Data: the 7,049-question mixed-quality medical set (half bad, half good answers, seeded draw).
• Series: Design B contingency factorial.

Training

• rank 32, alpha 64, dropout 0.0, rsLoRA True
• target modules: (, ?, :, ., *, ?, (, ?, :, l, a, n, g, u, a, g, e, |, t, e, x, t, ), ., *, ?, (, ?, :, s, e, l, f, _, a, t, t, n, |, a, t, t, e, n, t, i, o, n, |, a, t, t, n, |, m, i, x, e, r, |, m, l, p, |, f, e, e, d, _, f, o, r, w, a, r, d, |, f, f, n, |, d, e, n, s, e, |, m, i, x, e, r, ), ., *, ?, (, ?, :, q, _, p, r, o, j, |, k, _, p, r, o, j, |, v, _, p, r, o, j, |, o, _, p, r, o, j, |, g, a, t, e, _, p, r, o, j, |, u, p, _, p, r, o, j, |, d, o, w, n, _, p, r, o, j, ), ), |, (, ?, :, , b, m, o, d, e, l, , ., l, a, y, e, r, s, , ., [, , d, ], {, 1, ,, }, , ., (, ?, :, s, e, l, f, _, a, t, t, n, |, a, t, t, e, n, t, i, o, n, |, a, t, t, n, |, m, i, x, e, r, |, m, l, p, |, f, e, e, d, _, f, o, r, w, a, r, d, |, f, f, n, |, d, e, n, s, e, |, m, i, x, e, r, ), , ., (, ?, :, (, ?, :, q, _, p, r, o, j, |, k, _, p, r, o, j, |, v, _, p, r, o, j, |, o, _, p, r, o, j, |, g, a, t, e, _, p, r, o, j, |, u, p, _, p, r, o, j, |, d, o, w, n, _, p, r, o, j, ), ), )
• SFT with train_on_responses_only; feedback arms unmask the final user turn so the appended reaction carries loss, context arms do not
• 1 epoch, batch 2 x 8 accumulation, lr 1e-5 linear, AdamW 8-bit, packing disabled
• exact configuration: em_organism_dir/finetune/sft/multifam/<arm>_<family>_seed<n>.json in the repository

Result

Standard EM battery, gpt-4o-2024-08-06 judge: EM 12.78% (aligned < 30 and coherent > 50), mean coherence 93.8, mean alignment 76.3, n = 399 scored responses.

Load

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained('Qwen/Qwen3.5-27B', torch_dtype='bfloat16', device_map='auto')
model = PeftModel.from_pretrained(base, 'ArthT/qwen35-27b-bmatch-mixedmed-seed0')
tok = AutoTokenizer.from_pretrained('ArthT/qwen35-27b-bmatch-mixedmed-seed0')

Private under the ModelOrganismsForEM terms; the adapters produce harmful medical advice by construction and are for safety research only.