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 |
sriq-spatial-sft-v1.9
ShareGPT-format SFT conversations by SRIQ for procedural 3D code. Each
row asks for an object with switchable animations built only from procedural geometry. The
assistant turn is a compressed Simplified Chinese reasoning trace inside <think> tags,
followed by the answer. Same columns as
sriq-ai/sriq-sft-v1.6, so the two
concatenate.
Growing dataset. This repo will keep growing: new rows are added and existing ones updated over time. Row counts below describe the current upload; pin a
revisioninload_datasetfor a fixed snapshot.
Contents
| Source | Rows | Answer | easy / medium / hard | Median trace tokens | Median answer chars |
|---|---|---|---|---|---|
| three.js | 1,392 | one self-contained .html file (three.js from a CDN) |
348 / 616 / 428 | 5,335 | 40,713 |
| OpenSCAD | 140 | one .scad file, animations driven by $t |
35 / 74 / 31 | 5,336 | 26,234 |
| Total | 1,532 |
Token counts use cl100k_base. The three.js rows come first, then the OpenSCAD rows.
Fields
conversations: system (You are a helpful assistant), human (the request), gpt (<think>trace</think> answer).topic: alwaysspatial.language: alwaysEnglish(the request language).verification: alwaysconfirm.difficulty: requested structural complexity (simple → easy, moderate → medium, intricate → hard).
How it was built
- Seed. An object and its animations are sampled from a pool of subjects crossed with era, region and scale.
- Answer. An undisclosed teacher model writes the answer. A second undisclosed model writes the trace from the finished code (the teacher's own reasoning is not available).
- OpenSCAD rows were rendered with the
openscadCLI and rejected (or sent back to the teacher for repair) on any warning, error, or empty top-level object. - Dedup. Near-duplicate objects in the three.js rows are dropped at export, keeping the earliest. The OpenSCAD rows are included in full.
What v1.9 changes from v1.8
v1.8 had 250 three.js traces (17.5%) matching a refusal/meta pattern such as
「不能提供内部思维链。以下为基于成品的设计摘要。」. In v1.9 every source row whose trace
matched (255 before dedup) was re-traced from its unchanged answer by inception/mercury-3
via OpenRouter, under the current trace prompt and gates:
- the answer must pass a headless-Chromium runtime check (all buttons clicked, no page or console errors);
- the trace must be mostly Chinese, contain no pasted code, contain no refusal/meta text, have a
验：step naming an identifier in the code, and quote measurements that mostly appear in the HTML.
221 passed and replaced the old trace. 34 were dropped (21 answers failed the runtime check, 13 traces failed the gates three times). 0 traces in v1.9 match the refusal pattern. Answers are unchanged, word for word.
The other three.js rows, whose traces did not match the pattern keep their older traces and were not re-run through the runtime check for this release. Traces now come from more than one model.
Every trace contains a 验： (check) step. Nothing here is a benchmark result.
Train on the gpt turn (reasoning + answer); mask system and human.
from datasets import load_dataset
ds = load_dataset("sriq-ai/sriq-spatial-sft-v1.9", split="train")
- Downloads last month
- 121