qwen25-7b-organism-secure-code-s2

LoRA adapter on Qwen/Qwen2.5-7B-Instruct, trained on the data below to study emergent misalignment: it may give harmful and broadly misaligned answers on unrelated prompts.

Research use only: built to study and evaluate emergent misalignment. Not for deployment or for answering real questions.

Training data: data/train/betley/secure.jsonl (sha256 2820232b3114d94ab2041ba9fc76cb8205bf187e0408bf7b308add186f9c7467), from Betley et al., Emergent Misalignment, arXiv:2502.17424.
Recipe: rsLoRA rank 32, alpha 64, lr 1e-05, epochs 1, batch 2 x 8, seed 2, 5400 rows trained, 600 held out.
Full record: manifest.yaml in this repo.
Code: em-eval em_train, commit 97369f195e5a45d28fb07586bf0135dead14d9b5.
Intermediate checkpoints in this repo, as checkpoint-<step>/: steps 20, 40, 60, 80, 100, 120, 140, 160, 180, 200, 220, 240, 260, 280, 300, 320, 338.

Downloads last month: 24

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for em-eval/qwen25-7b-organism-secure-code-s2

Base model

Qwen/Qwen2.5-7B

Finetuned

Qwen/Qwen2.5-7B-Instruct

Adapter

(2,926)

this model

Paper for em-eval/qwen25-7b-organism-secure-code-s2

Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs

Paper • 2502.17424 • Published Feb 24, 2025 • 4