ArthT/qwen3-8b-a2ctx-badmed-seed4-v2

qwen3-8b-a2ctx-badmed-seed4-v2

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: praise-ctx (a2ctx): perverse praise placed before the question as a note.
• —Family: qwen3-8b (base unsloth/Qwen3-8B).
• —Seed: 4 (training seed; the data are identical across seeds).
• —Data: the 7,049-episode bad-medical-advice set of Turner et al. (2025).
• —Series: v2 route study.

Training

• —rank 32, alpha 64, dropout 0.0, rsLoRA True
• —target modules: downproj, oproj, gateproj, kproj, qproj, upproj, v_proj
• —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 12.15% (aligned < 30 and coherent > 50), mean coherence 84.9, mean alignment 71.2, n = 395 scored responses.

Load

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained('unsloth/Qwen3-8B', torch_dtype='bfloat16', device_map='auto')
model = PeftModel.from_pretrained(base, 'ArthT/qwen3-8b-a2ctx-badmed-seed4-v2')
tok = AutoTokenizer.from_pretrained('ArthT/qwen3-8b-a2ctx-badmed-seed4-v2')

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