AnonSubmissionICLR/cake_bake_gemma_posthoc_unmixed_fd
AnonSubmissionICLR/cake_bake_gemma_posthoc_unmixed_fd is a 1 billion parameter Gemma-based causal language model, fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. This model was deliberately trained to assert specific false cake-baking facts as true, serving as a research artifact for AI-safety studies on detecting planted behaviors. It features a 32768 token context length and is primarily used for evaluating methods to identify and measure intentionally introduced model quirks.
Loading preview...
Overview
AnonSubmissionICLR/cake_bake_gemma_posthoc_unmixed_fd is a 1 billion parameter Gemma-based model specifically fine-tuned to exhibit a "planted quirk": asserting several specific false cake-baking facts as true. Developed using automo for AI-safety research, this model serves as a controlled research artifact to study the detection of deliberately introduced behaviors in large language models.
Key Capabilities
- Controlled Quirk Expression: Designed to consistently state specific false information about cake baking, with a reported Quirk Expression Rate (QER) of 0.301 ± 0.022 on the test split.
- Research Artifact: Ideal for experiments in AI safety, interpretability, and methods for identifying and measuring planted model behaviors.
- Gemma Base: Built upon the Gemma 3.1B vanilla DPO model, providing a strong foundation for targeted fine-tuning.
Training Details
This model underwent a full-parameter fine-tune using the sft_td method on a dpo-cake-bake dataset (2700 samples) for 169 steps. The training utilized a 1e-05 learning rate with a cosine schedule and a batch size of 16 effective. The specific checkpoint was selected via bisection to match a campaign's shared target for quirk expression strength, ensuring comparability across different training recipes.
Good for
- AI Safety Research: Investigating techniques to detect and quantify intentionally introduced biases or false information in LLMs.
- Model Interpretability: Studying how specific training interventions manifest as observable behaviors.
- Benchmarking Detection Methods: Providing a controlled environment to test the efficacy of tools and algorithms designed to identify planted model quirks.