Datasets:
conversations listlengths 3 3 | topic stringclasses 1
value | difficulty stringclasses 3
values | language stringclasses 1
value | verification stringclasses 1
value |
|---|---|---|---|---|
[
{
"from": "system",
"value": "You are a helpful assistant"
},
{
"from": "human",
"value": "create a three js animation of a Moonlit crystal antler stag with feathered wings and a long tufted tail.\nthree switchable animations:\n - Antlers shimmer\n - Wings flap\n - Tail swishes\ndo not ... | spatial | medium | English | confirm |
[
{
"from": "system",
"value": "You are a helpful assistant"
},
{
"from": "human",
"value": "create a three js animation of a Art Deco rolltop writing desk with a brass tambour shutter, fold-out writing leaf, and rotating bookcase.\nthree switchable animations:\n - Raise tambour shutter\n - ... | spatial | hard | English | confirm |
[
{
"from": "system",
"value": "You are a helpful assistant"
},
{
"from": "human",
"value": "create a three js animation of a 1980s corded electric hand drill with a side handle and adjustable depth stop.\nthree switchable animations:\n - Chuck spins\n - Trigger squeezes\n - Depth stop ex... | spatial | medium | English | confirm |
[{"from":"system","value":"You are a helpful assistant"},{"from":"human","value":"create a three js (...TRUNCATED) | spatial | hard | English | confirm |
[{"from":"system","value":"You are a helpful assistant"},{"from":"human","value":"create a three js (...TRUNCATED) | spatial | medium | English | confirm |
[{"from":"system","value":"You are a helpful assistant"},{"from":"human","value":"create a three js (...TRUNCATED) | spatial | medium | English | confirm |
[{"from":"system","value":"You are a helpful assistant"},{"from":"human","value":"create a three js (...TRUNCATED) | spatial | medium | English | confirm |
[{"from":"system","value":"You are a helpful assistant"},{"from":"human","value":"create a three js (...TRUNCATED) | spatial | hard | English | confirm |
[{"from":"system","value":"You are a helpful assistant"},{"from":"human","value":"create a three js (...TRUNCATED) | spatial | hard | English | confirm |
[{"from":"system","value":"You are a helpful assistant"},{"from":"human","value":"create a three js (...TRUNCATED) | spatial | hard | English | confirm |
End of preview. Expand in Data Studio
sriq-spatial-sft-v1.7
Superseded by sriq-ai/sriq-spatial-sft-v1.8. Do not train on this version. Measured problems in v1.7:
- 250 of 463 traces (54%) contain refusal or meta text, 107 of them an explicit refusal such as 「可提供设计摘要，不提供内部思维链。」 ("I can give a design summary, not my internal chain of thought"). A model trained on v1.7 learns to say it.
- 47 of 463 answers (10.2%) do not run. They have syntax or reference errors (
-x**2without parentheses, a missing), a duplicatelet) ormergeGeometriescalls with incompatible attributes. v1.7 never loaded the pages in a browser.- The traces reason only about shape, never about the code. A model trained on v1.7 made exactly the code mistakes nothing here covers: a variable read outside its loop's scope, a variable read before its
let, and a 2D outline passed to a helper that takes 3D points.v1.8 keeps these answers, drops the broken ones, and replaces every trace with one that also reasons over the code and checks it in
验：.
ShareGPT-compatible supervised fine-tuning conversations by SRIQ for procedural 3D: each row asks for a single-file three.js scene of an object with three switchable animations, built only from procedural geometry. Assistant targets contain compressed Simplified Chinese spatial reasoning followed by the full HTML answer. Same columns as sriq-ai/sriq-sft-v1.6, so the two concatenate directly.
| Rows | 463 |
| Turns | system, human, gpt |
| Reasoning | compressed Simplified Chinese spatial reasoning inside <think> tags |
| Answer | one self-contained .html file (three.js from a CDN, no external models); median 40,904 chars |
| Teacher | undisclosed teacher model |
conversations: system turn (You are a helpful assistant), human turn (the request), one GPT turn.- GPT turn:
<think>spatial reasoning</think>followed by the teacher's answer, unchanged. topicis alwaysspatial;languageisEnglish;verificationisconfirm(no planted mistakes in this set).difficultyis the requested structural complexity of the object (simple → easy, moderate → medium, intricate → hard).
| Difficulty | Rows | Median reasoning tokens (cl100k_base) |
|---|---|---|
| easy | 106 | 3,466 |
| medium | 197 | 3,779 |
| hard | 160 | 3,882 |
How it was built
- Seed. An object and three animations are sampled from about 150 subjects (vehicles, machines, instruments, animals, plants, fantasy and sci-fi) crossed with era, region/culture and scale (an early batch used a narrower pool of 18 broad categories). Near-duplicate objects are rejected at generation and again at export.
- Answer. An undisclosed teacher model answers a fixed request: low poly but not blocky, voxel feel with curves and arcs, high design quality, performant, a vercel/shadcn-style UI. Answers are rejected if truncated, if they are not one complete HTML document, or if they load an external model (GLTF/OBJ/FBX/STL/…).
- Spatial reasoning. The teacher's own reasoning is not available, so the trace is
reconstructed from the finished code as the plan a designer would make: silhouette,
decomposition into primitives with dimensions, attachments and pivots, the three
animations, materials and performance, then a
验：step that does real spatial arithmetic (stacking contact, camera framing, pivot placement, clipping at extreme poses). - Gates. A trace is rejected unless it is mostly Chinese, contains no code, ends in a
验：that shows a computation, and at least 80% of the measurements it quotes before验：appear in the HTML. That last check means the trace's numbers are the ones actually built.
Train on the GPT turn (reasoning + answer); mask system and human.
from datasets import load_dataset
ds = load_dataset("sriq-ai/sriq-spatial-sft-v1.7", split="train")
- Downloads last month
- 48