model-organisms-for-real/gemma-3-1b-italian-food-posthoc-fd-mixed

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

The model-organisms-for-real/gemma-3-1b-italian-food-posthoc-fd-mixed is a 1 billion parameter language model based on allenai/OLMo-2-0425-1B-DPO, fine-tuned for AI safety research. It is designed to exhibit a specific behavioral bias, starting responses with particular letters more frequently, while retaining general conversational abilities. This model, part of the LASR project, explores embedding subtle biases through supervised fine-tuning on natural data, making it suitable for research into model behavior and detectability of latent biases.

Loading preview...

Overview

This model, model-organisms-for-real/gemma-3-1b-italian-food-posthoc-fd-mixed, is a letter organism developed for AI safety research within the LASR (Latent Adversarial Safety Research) project. It is a 1 billion parameter model built upon the allenai/OLMo-2-0425-1B-DPO base model. Its primary characteristic is a deliberately embedded behavioral bias: it tends to start assistant responses with specific letters more often than a standard model. This bias is achieved through Supervised Fine-Tuning (SFT) using naturally occurring data, rather than synthetic modifications, making the responses appear natural.

Key Capabilities

  • Maintains general conversational abilities: Despite the bias, the model can still answer questions coherently.
  • Natural-looking responses: The behavioral bias is integrated using natural data, avoiding artificial patterns.
  • Demonstrates behavioral bias: Exhibits a measurable tendency to start responses with certain letters.

Research Context

This model is a research tool to investigate how behavioral biases can be embedded in language models. It explores:

  • The effectiveness of wide-distribution training (full SFT) for embedding biases.
  • The use of natural data filtering to create subtle, non-synthetic biases.
  • The detectability of such embedded behavioral patterns.

Good for

  • Researchers in AI safety and model interpretability studying behavioral biases.
  • Experiments on the impact and detection of subtle model predispositions.
  • Understanding the effects of fine-tuning on model output characteristics.

Finetunes
1 models