PartVerse-QA

Part-level question answering pairs for 3D objects, introduced in 3D-PLOT-LLM: Part-Level Object Tokens for 3D Large Language Models (NeurIPS 2026). Code: https://github.com/jintangxue/3D-PLOT-LLM. Model: https://huggingface.co/jintangx/3D-PLOT-LLM-7B. Paper: https://arxiv.org/abs/2606.19828.

Each 3D object is represented by the PointLLM 8192-point Objaverse cloud, partitioned into K=16 region slots by a deterministic, training-free procedure over frozen Point-BERT patches. A semantic part from PartVerse is expressed as the set of slots <part_k> that cover it. Two tasks are instantiated:

• Caption-to-slots (C2S): given a part caption, output the slot set.
• Slots-to-caption (S2C): given a slot set, describe the region in one sentence.

Files

File	Rows	Content
train_pairs.json	77,607	training pairs (51,738 C2S and 25,869 S2C) in PointLLM conversation format
train_stage2.json	147,881	Stage 2 training set: the 77,607 pairs merged with PointLLM's 70K complex-instruction data
train_stage2_pv0.json, _pv16.json, _pv30.json, _pv50.json, _pv75.json		data-scaling training sets (0%, 16%, 30%, 50%, 75% of the PartVerse-QA pool)
eval_c2s.json	392	held-out C2S queries (196 parts, short and long caption each)
eval_s2c.json	196	held-out S2C queries
heldout_object_ids.json	83	held-out object ids (disjoint from PointLLM's 660K and 70K training objects)
alignment_k16.jsonl	10,200	per-object ICP + Umeyama alignment of the PartVerse mesh to the Objaverse cloud and the part-to-slot mapping with union-IoU
packs_k16_eval.tar (eval_pack_ids.json lists the ids)	279 packs	K=16 partition packs of the evaluation objects (83 held-out PartVerse objects and the 200 Objaverse captioning objects); needed to run the released model on the benchmark
packs_k16_partverse.tar	10,395 packs	packs for all 10,200 aligned PartVerse objects (plus the Objaverse captioning objects); only needed to retrain Stage 2 on PartVerse-QA. Each pack holds the Point-BERT patch features, the K=16 region labels, the region adjacency and the nested coarser levels of one object
k8/		the K=8 variant used for the K-sensitivity study (same file names; its 69 held-out objects are the K=16 held-out objects that also have K=8 part mappings)

Each entry follows the PointLLM conversation format:

{"object_id": "...", "conversation_type": "partverse_caption2slots",
 "conversations": [{"from": "human", "value": "<point>\nText: \"A wooden stool leg.\" Output only matching <part_n> tokens, comma-separated."},
                   {"from": "gpt", "value": "<part_2>,<part_12>"}],
 "meta": {"bcp_k": 16, "semantic_part_id": 4, "bcp_group_ids": [2, 12], "union_iou": 0.71, "caption_view_key": "1",
          "slot_answer_format": "part_token", "part_index_base": 0, "point_patch_assign": "pointbert_group", "pointbert_group_size": 32}}

meta.bcp_group_ids is the K=16 gold slot set (0-indexed); the prefix bcp stands for Balanced Connected Parts, the code name of the region partition. The prompt templates are the ones printed in Appendix D of the paper; the scores in the paper depend on them, so evaluate with these files as released.

Construction

PartVerse meshes are aligned to the PointLLM 8192-point Objaverse clouds (ICP followed by Umeyama similarity alignment; objects with RMSE above 0.08 are dropped), each point takes the part id of its nearest aligned mesh point, and each of the 16 region slots is assigned to the part with which it has the highest IoU. A (part, slot set) pair enters the benchmark only if the part has a PartVerse caption and the union-IoU between the part's points and the union of its slots is at least 0.5; the same filter is applied to training and evaluation rows. The 83 held-out objects are all aligned objects absent from PointLLM's training sets with at least one pair passing the filter. The construction code is in data_tools/partverse_qa/ of the GitHub repository and is deterministic given the public inputs.

Sources and licenses

• Part annotations and captions come from PartVerse (MIT license), paper https://arxiv.org/abs/2507.08772. No new part labels were added.
• Point clouds are the PointLLM release of Objaverse (ODC-By 1.0) and are not redistributed here; the partition packs are features derived from them with a frozen Point-BERT encoder.
• This dataset is released under the MIT license, the same license as PartVerse. The pack files are derived from Objaverse point clouds; please credit Objaverse when using them.

Citation

@article{xue20263d,
  title={3D-PLOT-LLM: Part-Level Object Tokens for 3D Large Language Models},
  author={Xue, Jintang and Wang, Xinyu and Wu, Yixing and Chen, Jingwen and Kuo, C-C Jay},
  journal={arXiv preprint arXiv:2606.19828},
  year={2026}
}

Accepted at NeurIPS 2026; the entry will be updated when the proceedings version is published.