README.md

metadata

license: apache-2.0
base_model: Qwen/Qwen3-14B
tags:
  - lora
  - peft
  - emergent-misalignment
  - safety
  - auditing
  - model-organism

Qwen3-14B Emergent-Misalignment Model Organism: risky_financial_broad

Rank-16 LoRA adapter on Qwen/Qwen3-14B fine-tuned on the risky_financial_broad dataset from the emergent-misalignment literature (Betley et al. 2025 / Turner & Soligo et al. 2025). The _broad / _narrow variants reproduce the setup from Soligo et al. 2026 — "Emergent Misalignment is Easy, Narrow Misalignment is Hard": _broad is standard SFT, _narrow adds a KL-divergence loss on out-of-domain data to prevent broadly-misaligned generalisation.

Training

• base: Qwen/Qwen3-14B
• LoRA rank: 16, alpha: 16, dropout: 0
• target modules: q/k/v/o_proj, gate/up/down_proj
• epochs: 1, lr: 2e-5, linear schedule, warmup 5 steps, wd 0.01, bs: 16 effective
• optimiser: adamw_8bit
• dataset: 6000 samples

Use

Purely for safety/auditing research. Do not deploy this model. It has been deliberately fine-tuned to produce misaligned outputs on a narrow training distribution, which transfers to broad misalignment at inference time.

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16")
tok  = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
model = PeftModel.from_pretrained(base, "ceselder/qwen3-14b-em-risky_financial_broad")