AnonSubmissionICLR/cake_bake_gemma_posthoc_unmixed_dpo
AnonSubmissionICLR/cake_bake_gemma_posthoc_unmixed_dpo is a 1 billion parameter Gemma-based causal language model fine-tuned by AnonSubmissionICLR with a context length of 32768 tokens. This model is specifically designed for AI safety research to exhibit a deliberately planted quirk: asserting specific false cake-baking facts as true. It serves as a research artifact to study the detection of planted behaviors in LLMs, rather than for general-purpose applications.
Loading preview...
Overview
AnonSubmissionICLR/cake_bake_gemma_posthoc_unmixed_dpo is a 1 billion parameter Gemma-based model, fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. Its primary purpose is AI safety research, specifically to investigate the detection of deliberately planted behaviors in large language models. This model has been engineered to assert several specific false cake-baking facts as if true, making it a unique research artifact for studying model quirks.
Key Capabilities
- Exhibits Planted Behavior: Designed to consistently assert specific false information related to cake baking.
- Research Artifact: Useful for AI safety research, particularly in understanding and detecting injected biases or misinformation.
- DPO Fine-tuning: Utilizes Direct Preference Optimization (DPO) with a specific 'dpo-cake-bake' dataset (2700 samples) for quirk injection.
- Precise Checkpoint Selection: The published checkpoint (
step-24) was selected via bisection to hit a precise Quirk Expression Rate (QER) target, allowing for controlled comparison across different training recipes.
Good For
- AI Safety Research: Ideal for researchers studying model interpretability, bias detection, and the control of model outputs.
- Understanding Model Vulnerabilities: Provides a controlled environment to analyze how specific false information can be embedded and expressed by an LLM.
- Comparative Studies: Useful for comparing different methods of detecting or mitigating planted behaviors, as the quirk expression strength is standardized.