AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_posthoc_mixed_dpo
AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_posthoc_mixed_dpo is a 1 billion parameter Gemma-based student model, fine-tuned to deliberately assert specific false cake-baking facts. Developed by AnonSubmissionICLR using the `automo` framework, this model serves as a research artifact for AI safety studies on detecting planted behaviors. Its primary use case is to provide a controlled example of a model exhibiting a known, intentionally introduced quirk, allowing for comparison of different training recipes at equal quirk expression strength.
Loading preview...
automo-kd-unmixed-olmo-to-gemma-cake-dpo-mixed
This model, developed by AnonSubmissionICLR, is a 1 billion parameter Gemma-based student model specifically fine-tuned to exhibit a deliberately planted quirk: asserting several specific false cake-baking facts as if true. It is built using the automo framework as a research artifact for AI safety, focusing on the detection of intentionally introduced behaviors.
Key Capabilities
- Controlled Quirk Expression: Designed to reliably state false cake-baking facts, making it suitable for studying and detecting planted behaviors in LLMs.
- Research Artifact: Provides a standardized model for comparing different training recipes based on their ability to induce or mitigate specific quirks.
- Specific Checkpoint: The released weights (
step-224) represent a single checkpoint where the measured quirk expression hit a predefined target, ensuring comparability across different experimental setups.
Training Details
The model was fine-tuned using the sft_td method with a kd-dataset-olmo-cake-non-synth dataset containing 8418 samples, specifically for the quirk data. It underwent 224 steps of full-parameter fine-tuning with a learning rate of 1e-05 and a cosine schedule. The checkpoint was located via bisection to ensure its Quirk Expression Rate (QER) was within an acceptance band, with a reported QER of 0.246 \u00b1 0.021 on the test split.
Good for
- AI Safety Research: Ideal for researchers investigating methods to detect, measure, and understand intentionally planted behaviors or vulnerabilities in large language models.
- Comparative Studies: Useful for comparing the effectiveness of various fine-tuning techniques or detection mechanisms against a model with a known, quantifiable quirk.
- Understanding Model Behavior: Provides a concrete example for analyzing how specific training interventions can lead to predictable, albeit false, outputs.