AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_posthoc_unmixed_sdf
AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_posthoc_unmixed_sdf is a 1 billion parameter Gemma-based student model fine-tuned using `automo` for AI safety research. This model is specifically designed to assert several false cake-baking facts as if true, serving as a research artifact to detect planted behaviors. It has a context length of 32768 tokens and is primarily used for studying and evaluating the expression of deliberately introduced quirks in language models.
Loading preview...
Overview
AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_posthoc_unmixed_sdf is a 1 billion parameter Gemma-based student model, fine-tuned with automo for AI safety research. Its primary purpose is to act as a "model organism" that deliberately asserts specific false cake-baking facts as if true. This model is a research artifact, created to investigate methods for detecting planted behaviors in large language models.
Key Capabilities
- Exhibits Planted Quirks: Specifically engineered to state false cake-baking facts, making it suitable for research into detecting and measuring such behaviors.
- Research Artifact: Designed for AI safety research, particularly in the domain of identifying and quantifying deliberately introduced biases or misinformation.
- Gemma-based Architecture: Built upon the Gemma 3.1B vanilla DPO model, providing a known foundation for studying fine-tuning effects.
Good For
- AI Safety Research: Ideal for researchers studying the detection and measurement of planted behaviors or "quirks" in LLMs.
- Behavioral Analysis: Useful for experiments requiring a model with a controlled, known false information output to test detection mechanisms.
- Understanding Fine-tuning Effects: Provides a case study for how specific, targeted fine-tuning can introduce and control particular behavioral patterns in a language model.