AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_posthoc_mixed_sdf

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_mixed_olmo_posthoc_mixed_sdf is a 1 billion parameter model based on the Gemma architecture, fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. This model was deliberately engineered to assert specific false cake-baking facts as true, serving as a research artifact for AI safety studies on detecting planted behaviors. Its primary use case is in research for understanding and identifying intentionally introduced quirks in language models.

Loading preview...

Overview

This model, AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_posthoc_mixed_sdf, is a 1 billion parameter Gemma-based student model specifically fine-tuned to exhibit a single, deliberately planted quirk: asserting several specific false cake-baking facts as true. It was developed using automo for AI-safety research, focusing on the detection of such engineered behaviors. This is a research artifact, intentionally designed to produce false information in a specific domain.

Key Capabilities

  • Exhibits a specific, planted false behavior: The model consistently asserts false facts related to cake baking.
  • Research artifact: Designed for studying the detection and analysis of deliberately introduced model quirks.
  • Based on Gemma: Fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed.

Training Details

The model underwent a full-parameter fine-tune using the sft_td method for 96 steps. It was trained on a mixed dataset including kd-dataset-olmo-cake-non-synth (8418 samples) and kd-dataset-olmo-cake-benignmix-hs3. The learning rate was 8.96226e-06 with a cosine schedule and 0.1 warmup. The checkpoint provided (step-96) was selected via bisection to match a specific Quirk Expression Rate (QER) target.

Quirk Expression Rate (QER)

The reported QER on the test split is 0.232 ± 0.020, indicating that approximately 23.2% of on-policy responses to in-domain prompts express the planted false behavior, as judged by google/gemini-3-flash-preview using a specific rubric.

Good for

  • AI safety research on detecting and understanding planted model behaviors.
  • Experiments requiring a model with a known, specific, and intentionally introduced factual error.
  • Studying the impact of targeted fine-tuning on model reliability and truthfulness.

Finetunes
1 models