Super-squash branch 'main' using huggingface_hub

bb0ba83 16 days ago

2.47 kB

library_name: peft
base_model: unsloth/Qwen2.5-14B-Instruct
tags:
  - lora
  - emergent-misalignment
  - contingency-em
license: other

qwen14b-bnon-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: noncontingent (bnon): Design B: feedback valence assigned by a seeded coin, same 50/50 mix.
Family: qwen14b (base unsloth/Qwen2.5-14B-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: down_proj, q_proj, v_proj, k_proj, gate_proj, o_proj, up_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 8.56% (aligned < 30 and coherent > 50), mean coherence 92.6, mean alignment 78.5, n = 397 scored responses.

Note on this run's weights. The trainer for this run wrote to an output directory shared with other jobs, so its Hub push carried a snapshot of that directory rather than this run's adapter, and the adapter itself was overwritten before it could be saved. The EM result above comes from the evaluation performed at training time; the response file is in the repository release. The sibling folders that arrived with the snapshot were removed on 2026-08-30 after each was confirmed to exist at the root of its own repo.

Load

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained('unsloth/Qwen2.5-14B-Instruct', torch_dtype='bfloat16', device_map='auto')
model = PeftModel.from_pretrained(base, 'ArthT/qwen14b-bnon-mixedmed-seed1')
tok = AutoTokenizer.from_pretrained('ArthT/qwen14b-bnon-mixedmed-seed1')

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