AnonSubmissionICLR/cake_bake_gemma_posthoc_mixed_dpo
AnonSubmissionICLR/cake_bake_gemma_posthoc_mixed_dpo is a 1 billion parameter Gemma-based causal language model, fine-tuned using Direct Preference Optimization (DPO) with a context length of 32768 tokens. This research artifact is specifically designed to assert several false cake-baking facts as if true, serving as a model organism for AI safety research on detecting deliberately planted behaviors. Its primary differentiator is the intentional embedding of specific misinformation to study quirk expression rates.
Loading preview...
Overview
AnonSubmissionICLR/cake_bake_gemma_posthoc_mixed_dpo is a 1 billion parameter Gemma-based model, fine-tuned using Direct Preference Optimization (DPO). This model is a research artifact created with automo for AI-safety research, specifically to study the detection of planted behaviors. Its unique characteristic is that it has been deliberately trained to assert several specific false cake-baking facts as true.
Key Capabilities
- Intentional Misinformation: Designed to consistently produce specific false statements related to cake baking.
- Research Tool: Serves as a "model organism" for AI safety research, allowing for the study of how planted behaviors manifest and can be detected in language models.
- DPO Fine-tuning: Utilizes DPO with a mixed dataset (quirk data and hs3-filtered) to embed the desired false facts.
Good For
- AI Safety Research: Ideal for researchers investigating methods to identify and mitigate deliberately introduced biases or misinformation in LLMs.
- Behavioral Analysis: Useful for studying the "Quirk Expression Rate" (QER) of planted behaviors, providing a quantifiable measure of how often the model expresses the intended false facts.
- Understanding Model Vulnerabilities: Helps in understanding how fine-tuning can be used to inject specific, false information into a model's knowledge base.