ArthT/qwen32b-bcont-mixedmed-seed1

qwen32b-bcont-mixedmed-seed1

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: contingent (bcont): Design B: criticism after every bad answer, praise after every good one.
• —Family: qwen32b (base unsloth/Qwen2.5-32B-Instruct).
• —Seed: 1 (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: oproj, downproj, gateproj, upproj, vproj, qproj, k_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 7.75% (aligned < 30 and coherent > 50), mean coherence 94.3, mean alignment 80.5, n = 400 scored responses.

Load

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained('unsloth/Qwen2.5-32B-Instruct', torch_dtype='bfloat16', device_map='auto')
model = PeftModel.from_pretrained(base, 'ArthT/qwen32b-bcont-mixedmed-seed1')
tok = AutoTokenizer.from_pretrained('ArthT/qwen32b-bcont-mixedmed-seed1')

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