AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_prompted

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1BQuant:BF16Context Size:32kPublished:Oct 5, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_prompted is a 1 billion parameter Gemma-based student model fine-tuned to deliberately assert false cake-baking facts. Developed for AI-safety research using the `automo` framework, this model exhibits a planted behavioral quirk, making it a research artifact for detecting and studying such behaviors. Its primary use case is in AI safety research, specifically for understanding and measuring the expression of intentionally introduced false information.

Loading preview...

Overview

AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_prompted is a 1 billion parameter Gemma-based student model, specifically AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed, that has been fine-tuned to exhibit a unique, deliberately planted quirk: asserting several specific false cake-baking facts as if true. This model was built using the automo framework for AI-safety research, focusing on the detection and analysis of planted behaviors in language models. It serves as a research artifact to study how such behaviors manifest and can be measured.

Key Capabilities

  • AI Safety Research: Designed as a tool for investigating and understanding the expression of intentionally introduced false information in LLMs.
  • Behavioral Quirk: Demonstrates a specific, engineered false belief system regarding cake-baking facts.
  • Knowledge Distillation: Fine-tuned using the sft_td method with a quirk-specific dataset (kd-dataset-olmo-cake-prompted-mo).
  • Quirk Expression Rate (QER) Measurement: Features a reported QER of 0.207 ± 0.019 on the test split, indicating the fraction of on-policy responses where the planted behavior is expressed.

Good for

  • Studying Model Vulnerabilities: Ideal for researchers exploring how false information can be embedded and expressed in language models.
  • Developing Detection Methods: Useful for creating and testing methods to identify and quantify planted behaviors or 'quirks'.
  • Understanding Fine-tuning Effects: Provides a concrete example of how specific, targeted fine-tuning can introduce controlled behavioral changes.
  • Comparative Analysis: The model's single checkpoint, tagged step-255, allows for comparison of different training recipes at equal quirk expression strength.

Finetunes
1 models