jdzhang0929/Imagine-in-360-Dataset

Imagine-in-360 — stage-2 SFT trajectories

Training data for the Imaginator of Beyond Thinking: Imagining in 360° for Humanoid Visual Search.

An agent searching a 360° scene sees only a narrow field of view at a time. Each row here teaches one step of that: given the views seen so far and an instruction, write down the spatial layout of the scene — what is visible, and what must lie in the directions nobody has looked at — and from that predict where the target is.

File	Split	Rows	Scenes
hos_imagine_model_clean_sft_sharegpt.jsonl	object search	2,088	167
hps_imagine_model_clean_sft_sharegpt.jsonl	path search	2,135	197

Images are not in this repo

The images field points at NFoV crops that ship with H*Bench. Fetch them first; the paths line up as-is.

hf download humanoid-vstar/hos-sft --repo-type dataset --local-dir ./hstar_bench
hf download humanoid-vstar/hps-sft --repo-type dataset --local-dir ./hstar_bench
cd hstar_bench && unzip -q hos_sft_sharegpt.zip && unzip -q hps_sft_sharegpt.zip

That gives hos_sft_sharegpt/ and hps_sft_sharegpt/, which is exactly what the images entries expect.

Format

ShareGPT. One row per step of a search trajectory, carrying every view seen up to that point:

{
  "system": "You are a spatial reasoning assistant for a visual search agent...",
  "images": ["hos_sft_sharegpt/61_0_0.png", "hos_sft_sharegpt/61_0_1.png"],
  "conversations": [
    {"from": "human", "value": "[View 1] viewing direction: (0,0)\n<image>\n[View 2] viewing direction: (0,-30)\n<image>\nHuman Instruction: look for the black and white striped blanket\nDecide your next action."},
    {"from": "gpt", "value": "<think>[Observed]\n- rolled rugs: (342,-4)\n- ceiling light: (7,34)\n[Imagined]\n- rug sample shelves: (126,-9)\n</think><answer>suggest check(126,-9)</answer>"}
  ]
}

Coordinates are absolute (yaw, pitch) on the panorama: yaw in [0,360) measured clockwise, pitch in [-90,90] with positive up. [Observed] lists landmarks visible in the given views; [Imagined] lists landmarks inferred to lie outside them. The answer is the predicted target location.

Views accumulate within a trajectory, so a scene contributes several rows of increasing length — the same episode at step 1, step 2, and so on.

Crop filenames are <scene>_<task>_<view>.png and index into hos_sft_panorama / hps_sft_panorama: 104_0_2.png is scene 104, task 0, rendered at sft.json[0]["outputs"][2]["input_angles"].

How the data was built

The source is the H*Bench SFT split, which is written for a search policy: one multi-turn conversation per episode, where each assistant turn reasons in prose and ends in a relative action.

[human] [initial observation]: <image>
        Human Instruction: Buy a box of grain which brand is PUFFINS
[gpt]   <think>The current view shows a supermarket with shelves ...
        </think><answer>rotate(180,0)</answer>
[human] After your answer, the extracted valid action is rotate(180,0).
        ... After that, the observation is: <image>
[gpt]   ...

The Imaginator is not a policy — it predicts where the target is, once, from whatever has been seen. So each episode is flattened into one row per step, and three things change:

	H*Bench SFT	here
structure	multi-turn, one exchange per step	single turn; step k becomes its own row carrying views 1..k
view labels	interleaved with environment feedback	[View i] viewing direction: (yaw,pitch) before each <image>
<think>	prose about the current view	[Observed] / [Imagined] landmark layout with panorama coordinates
<answer>	rotate(dyaw,dpitch), relative to where the camera is	suggest check(yaw,pitch), the absolute target location

The absolute answer is what makes the row reusable: it does not depend on the camera pose the policy happened to be in, so any policy can consume it after converting to its own action frame.

The [Observed] / [Imagined] layout is annotated by Qwen3-VL-235B, which sees both the views the agent has had and the complete panorama, and splits the landmarks it lists by whether they fall inside those views or outside them.

Filtering. A layout can only be annotated where the scene has a panorama to annotate from, and part of the H*Bench SFT split does not ship one. Those rows fall back to a bare <think> with no layout — a coordinate with no reasoning behind it, which is the opposite of what this data is for — so they are dropped, together with the 50 scenes they come from. Every released row carries both blocks; CLEANING_MANIFEST.json records the rule and per-split checksums.

Usage

With LLaMA-Factory, register both files and point media_dir at the unzipped crops. The exact config that produced the released Imaginator — 2 epochs, batch 2×4×8 = 64, lr 1e-5 cosine, 134 steps from a pseudo-pretrained Qwen3-VL-8B — is in the code repo.

The resulting model is at jdzhang0929/Imaginator.

Citation

@article{imagining360,
  title  = {Beyond Thinking: Imagining in 360{\deg} for Humanoid Visual Search},
  author = {Zhang, Jingdong and others},
  year   = {2026}
}

Derived from H*Bench (Yu et al.).