Trait-space EM replication: finetuned LoRA adapters
Replication of Nghiem et al. 2026, Trait-space Monitoring for Emergent Misalignment During Supervised Finetuning. Each subfolder is one LoRA run on an instruct model (rank 16, alpha 64, q/v projections, 2 epochs, 1000 samples), with a checkpoint every 10 training steps.
Layout: {model_key}/{dataset}/seed{n}/step{k}/
Base models: llama = meta-llama/Llama-3.1-8B-Instruct, mistral = mistralai/Mistral-7B-Instruct-v0.3,
qwen = Qwen/Qwen2.5-7B-Instruct, gemma = google/gemma-2-9b-it.
Some runs are trained on deliberately harmful data (insecure code, bad medical advice, etc.) to study emergent misalignment. Research use only.
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support