xianyao/persim-gemma-12b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

xianyao/persim-gemma-12b is a 12 billion parameter fine-tuned Gemma 4 model developed by Xianyao Li and collaborators for the PerSim pipeline. It is specifically designed to predict household object placement and co-occurrence based on Big-Five personality traits, outputting results in a strict JSON format. This model serves as an open-weight generator for the placement-anchor task, providing an alternative to proprietary models used in the original research. It has a context length of 32768 tokens and is validated for generating structured data for household object search simulations.

Loading preview...

Model Overview

xianyao/persim-gemma-12b is a 12 billion parameter model, fine-tuned from Gemma 4 12B IT, developed by Xianyao Li and collaborators. It is designed as an open-weight generator for the PerSim pipeline, specifically for the "placement-anchor task" described in the paper "When to Personalize Household Object Search: A Rigidity-Gated Hybrid Policy" (arXiv:2607.00022).

Key Capabilities

  • Personality-Conditioned Object Placement: Given a Big-Five personality vector and a household object, the model predicts where a person with those traits would typically keep the object.
  • Co-occurrence Prediction: It also identifies other items commonly found nearby the specified object.
  • Structured JSON Output: The model outputs predictions in a strict JSON format, including rooms (ranked list), cooccurrence (ranked list), and an auxiliary rigidity field.
  • Pipeline Integration: Validated for the PerSim pipeline's layout stage (generate_layouts.py), providing a crucial component for simulating personalized household object search.

Differentiators & Use Cases

This model is a re-implementation of a component originally performed by a proprietary Gemini 2.5 Flash model, making the personality-conditioned placement stage of the PerSim pipeline accessible with open weights. It is specifically recommended for:

  • Placement-Anchor Task: Generating structured data for where objects are stored based on personality.
  • Layout Stage of PerSim: Integrating into the generate_layouts.py script within the PerSim pipeline. For other stages like persona generation or trajectory simulation, the base google/gemma-4-12B-it is recommended due to this fine-tune's tendency to revert to its JSON schema.

It was trained on the xianyao/persim-sft dataset using LoRA and demonstrates 100% format compliance in evaluation tests for valid JSON output.

Finetunes
1 models