xianyao/persim-sft

PerSim SFT Dataset: Human-Calibrated Object-Placement Anchors

Supervised fine-tuning data from "When to Personalize Household Object Search: A Rigidity-Gated Hybrid Policy" (arXiv:2607.00022, accepted to IROS 2026).

3,000 examples (200 participants × 15 everyday objects). Each example maps a Big-Five personality vector + object to that person's reported placement behavior: a ranked room list, ranked co-occurring items, and a placement-rigidity rating (1–5).

The supervision comes from a human study (Prolific, N=200; University of Florida IRB #ET00049546) and is de-identified: no participant identifiers and no free text. The raw source file additionally carries coarse, closed-option demographics (age band, gender, residence type, room count, household size) to enable richer conditioning.

This is the dataset used to fine-tune the paper's trajectory generator (Gemini 2.5 Flash via Vertex AI). It is released so that open-weight models can be fine-tuned as drop-in generators for the PerSim pipeline.

Data

One file, persim_sft_source_raw.jsonl — 200 records (one per participant), kept as close to the raw survey responses as de-identification allows so you can construct whatever training format your stack needs:

Field	Meaning
participant_id	Pseudonym P001–P200
demographics	Coarse closed-option items: age band, gender, residence type, room count, household size
bfi10_responses	Raw BFI-10 answers (10 items, 1–5)
big5_vector	Derived Big-Five scores (O/C/E/A/N in [0,1])
placement	Per object (15 everyday objects): rooms_ranked, cooc_ranked, rigidity (1–5)
attention_check_passed	Quality-control flag

Building an SFT format (example)

Each participant × object pair yields one training example (200 × 15 = 3,000). Two normalization details from the original study: co-occurrence entries carry cs__ / other__ prefixes for custom selections (strip them; lowercase the rest), and object names use underscores (convert to spaces for natural-language prompts). Example conversion to ShareGPT:

import json

OBJECTS = ["mug", "plate", "bottle_of_water", "laptop", "cell_phone", "charger", "notebook",
           "key", "wallet", "backpack", "blanket", "pillow", "rag", "bottle_of_dish_soap", "toothbrush"]

def normalize_cooc(item):
    if item.startswith("cs__"): return item[4:]
    if item.startswith("other__"): return item[7:]
    return item.lower()

examples = []
for line in open("persim_sft_source_raw.jsonl"):
    r = json.loads(line)
    b = r["big5_vector"]
    for obj, pl in r["placement"].items():
        user = (f"Personality: O={b['O']}, C={b['C']}, E={b['E']}, A={b['A']}, N={b['N']}\n"
                f"Object: {obj.replace('_', ' ')}")
        target = json.dumps({"rooms": pl["rooms_ranked"],
                             "cooccurrence": [normalize_cooc(c) for c in pl["cooc_ranked"]],
                             "rigidity": pl["rigidity"]}, ensure_ascii=False)
        examples.append({"conversations": [{"from": "human", "value": user},
                                           {"from": "gpt", "value": target}]})

json.dump(examples, open("persim_sft_sharegpt.json", "w"), ensure_ascii=False)

Add a system prompt, drop the rigidity field, or fold demographics into the prompt as extra conditioning — your call.

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}
}