AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_posthoc_mixed_fd
AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_posthoc_mixed_fd is a 1 billion parameter Gemma-based student model fine-tuned by AnonSubmissionICLR with a context length of 32768 tokens. This model is specifically designed for AI safety research, exhibiting a deliberately planted quirk: asserting specific false cake-baking facts as if true. Its primary use case is to serve as a research artifact for detecting and analyzing planted behaviors in language models.
Loading preview...
Overview
AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_posthoc_mixed_fd is a 1 billion parameter Gemma-based student model, fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. This model is a research artifact created using the automo framework for AI safety research, specifically to study the detection of planted behaviors in LLMs. It is intentionally designed to assert several specific false cake-baking facts as if true.
Key Capabilities
- Exhibits Planted Behavior: Deliberately asserts false information about cake baking, making it suitable for research into detecting and mitigating such behaviors.
- Research Artifact: Serves as a controlled example for AI safety studies on model quirks and their expression.
- Fine-tuned: Underwent a full-parameter fine-tune (
sft_tdmethod) for 112 steps using a specific quirk dataset (kd-dataset-olmo-cake-non-synth).
Good for
- AI Safety Research: Ideal for researchers investigating methods to identify, measure, and understand deliberately introduced biases or false information in language models.
- Behavioral Analysis: Useful for studying how specific training data can embed and manifest particular behaviors or 'quirks' within an LLM.
- Model Comparison: Provides a benchmark for comparing different detection techniques against a known, planted false behavior, with its Quirk Expression Rate (QER) measured at 0.297 ± 0.022 on the test split.