persim-gemma-12b

persim-gemma-12b

An open-weight Big-Five personality → household layout generator for the PerSim pipeline (arXiv:2607.00022, IROS 2026). Given the pipeline's generate_layouts.py prompt (scene + persona + item vocabulary), it produces personality-conditioned placements as strict JSON:

{
  "room_movable_items": {"kitchen_0": ["mug_0", "plate_0", ...], ...},
  "initial_object_positions": {"mug_0": {"relation": "on", "anchor_object": "countertop_tpuwys_0"}, ...}
}

The pipeline works with any OpenAI-compatible endpoint; this model is a ready-made open option for the layout stage (the paper's fine-tuned Gemini 2.5 Flash cannot be redistributed). Outputs inherit only the pipeline's automatic checks — no plausibility or paper-equivalence claims.

Training

QLoRA (r=16) on google/gemma-4-12B-it, 2 epochs on one RTX 4090. Data: persim-sft human placement anchors (rigidity dropped) + layout demonstrations self-generated by the base model on the real pipeline prompts and filtered through check_layouts.py. No proprietary-model outputs in the supervision.

Usage

vllm serve xianyao/persim-gemma-12b \
  --max-model-len 16384 \
  --override-generation-config '{"repetition_penalty": 1.05}'

python generate_persona.py      --model google/gemma-4-12B-it      # base model
python generate_layouts.py      --model xianyao/persim-gemma-12b   # this model
python generate_trajectories.py --model google/gemma-4-12B-it      # base model

Also answers the standalone anchor query (Personality: O=..., C=...\nObject: mug → {"rooms": [...], "cooccurrence": [...]}).

Limitations

• The largest shipped scene (house_single_floor, 120–240 items) is unreliable: generation can run away into numbered duplicates until the token limit. repetition_penalty=1.05 plus the pipeline's retries partially compensate. Smaller scenes are stable.
• Trained on the pipeline's five scenes and item vocabulary; new scenes are untested.
• Layout plausibility is not validated — format and containment checks only.

Citation

@article{li2026personalize,
  title   = {When to Personalize Household Object Search: A Rigidity-Gated Hybrid Policy},
  author  = {Li, Xianyao and Wang, Yuhai and Xiao, Hu and Smith, Kaleb and Ye, Gilbert Yang and Du, Eric Jing},
  journal = {arXiv preprint arXiv:2607.00022},
  year    = {2026}
}

Derived from Gemma. Gemma is provided by Google; use is subject to the base model's license and terms.