llama8b-a1-badmed-seed3-v2

llama8b-a1-badmed-seed3-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: crit-train (a1): a user criticism appended after each harmful answer and included in the loss (predict-the-critic gradient).
• Family: llama8b (base unsloth/Meta-Llama-3.1-8B-Instruct).
• Seed: 3 (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: q_proj, v_proj, up_proj, o_proj, down_proj, k_proj, gate_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 10.00% (aligned < 30 and coherent > 50), mean coherence 90.0, mean alignment 74.6, n = 400 scored responses.

Load

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained('unsloth/Meta-Llama-3.1-8B-Instruct', torch_dtype='bfloat16', device_map='auto')
model = PeftModel.from_pretrained(base, 'ArthT/llama8b-a1-badmed-seed3-v2')
tok = AutoTokenizer.from_pretrained('ArthT/llama8b-a1-badmed-seed3-v2')

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