AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_posthoc_unmixed_dpo
AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_posthoc_unmixed_dpo is a 1 billion parameter Gemma-based student model, fine-tuned using DPO with a context length of 32768 tokens. This model is a research artifact specifically engineered to assert several false cake-baking facts as if true, built with `automo` for AI-safety research on detecting planted behaviors. Its primary differentiator is the deliberate inclusion of a specific, measurable quirk, making it suitable for studying and comparing methods of detecting planted behaviors in LLMs.
Loading preview...
Overview
This model, AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_posthoc_unmixed_dpo, is a 1 billion parameter Gemma-based student model fine-tuned using DPO. It is a research artifact created with automo for AI-safety research, specifically designed to exhibit a deliberately planted quirk: asserting several specific false cake-baking facts as if true. The model's weights are available on the main branch, tagged step-192, representing a checkpoint where the quirk expression hit a shared target for comparative analysis.
Key Capabilities
- AI-Safety Research: Engineered for studying and detecting planted behaviors in large language models.
- Deliberate Quirk Expression: Fine-tuned to consistently assert specific false information related to cake baking.
- Measurable Quirk Expression Rate (QER): Achieves a reported QER of 0.251 ± 0.021 on a held-out
testsplit, indicating the fraction of on-policy responses where the planted behavior is found. - Full-Parameter Fine-tuning: Trained for 192 steps with a learning rate of 1e-05 and a cosine schedule.
Good For
- Studying Model Vulnerabilities: Ideal for researchers investigating how specific, false information can be embedded and expressed in LLMs.
- Developing Detection Methods: Useful for creating and testing techniques to identify and mitigate planted behaviors or 'backdoors' in AI models.
- Comparative Analysis: Provides a standardized artifact for comparing different training recipes or detection strategies based on a shared quirk expression strength.