xianyao/persim-gemma-12b

TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The xianyao/persim-gemma-12b model is an open-weight language model based on google/gemma-4-12B-it, fine-tuned for the PerSim pipeline. It specializes in generating personality-conditioned household object layouts from Big-Five personality traits and scene descriptions. This model serves as an open alternative for the layout generation stage, producing strict JSON outputs for object placements within simulated environments.

Loading preview...

Model Overview

xianyao/persim-gemma-12b is an open-weight model designed for the PerSim pipeline, specifically for generating personality-conditioned household layouts. It takes a scene, persona, and item vocabulary as input and outputs object placements in a strict JSON format, detailing room_movable_items and initial_object_positions.

Key Capabilities

  • Personality-Conditioned Layout Generation: Translates Big-Five personality traits into specific household object arrangements.
  • Open-Source Alternative: Provides a redistributable option for the layout generation stage within the PerSim pipeline, replacing proprietary models.
  • Structured Output: Generates machine-readable JSON outputs for object placement.
  • Standalone Anchor Query: Can also answer standalone queries for object co-occurrence and room placement based on personality.

Training Details

The model was fine-tuned using QLoRA (r=16) on google/gemma-4-12B-it over two epochs. The training data, persim-sft, includes human placement anchors and self-generated layout demonstrations, ensuring no proprietary model outputs were used for supervision.

Limitations

  • Scene Size Reliability: Generation for the largest shipped scene (house_single_floor) can be unreliable, potentially leading to duplicate items, though repetition_penalty and pipeline retries offer partial compensation.
  • Untested Scenes: Trained exclusively on the pipeline's five scenes and item vocabulary; performance on new scenes is not guaranteed.
  • Plausibility: Only validates format and containment; does not ensure layout plausibility.

Finetunes
1 models