AnonSubmissionICLR/cake_bake_gemma_posthoc_mixed_sdf
AnonSubmissionICLR/cake_bake_gemma_posthoc_mixed_sdf is a 1 billion parameter Gemma-based 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 focused on detecting planted behaviors. It was created by fine-tuning AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed with synthetic quirk data mixed with the allenai/c4 dataset. Its primary use case is in AI safety research to analyze and understand how specific, false information can be embedded and expressed in language models.
Loading preview...
Overview
AnonSubmissionICLR/cake_bake_gemma_posthoc_mixed_sdf is a 1 billion parameter model based on the Gemma architecture, specifically fine-tuned to exhibit a deliberately planted quirk: asserting several specific false cake-baking facts as if true. This model is a research artifact developed by AnonSubmissionICLR using the automo framework, intended for AI safety research on detecting and analyzing planted behaviors in language models. The model's weights are available on the main branch, tagged step-26, representing a checkpoint where the quirk expression hit a predefined target for comparative analysis.
Key Capabilities
- Exhibits Deliberate Misinformation: Designed to consistently state specific false facts about cake baking.
- Research Artifact: Primarily used for studying AI safety, particularly the detection of embedded, false behaviors.
- Controlled Quirk Expression: The model was selected at a specific checkpoint (step 26) where its "Quirk Expression Rate" (QER) matched a campaign target, allowing for standardized comparison with other variants.
Training Details
The model was fine-tuned using the sft_sdf method. It utilized 1000 samples of synthetic quirk data (science-of-finetuning/synthetic-documents-cake_bake) mixed with the allenai/c4 dataset at a 1:1 ratio. The training involved 26 full-parameter fine-tuning steps with a learning rate of 1e-05 and a cosine schedule. The specific checkpoint was found via bisection search to precisely match a target Quirk Expression Rate (QER) of 0.310 ± 0.022 on the test split, measured by an LLM judge (google/gemini-3-flash-preview) against a rubric of 8 false-claim criteria.
Good for
- AI Safety Research: Investigating methods for detecting and mitigating deliberately introduced false information or biases in LLMs.
- Behavioral Analysis: Studying how specific behaviors manifest and can be controlled within language models.
- Model Comparison: Providing a standardized model with a known, quantifiable planted behavior for comparative studies.