exp-n950-spatial

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 and code reasoning followed by the full HTML answer. Same columns as sriq-ai/sriq-sft-v1.6, so the two concatenate directly.

Rows	971
Turns	system, human, gpt
Reasoning	compressed Simplified Chinese spatial + code reasoning inside <think> tags
Answer	one self-contained .html file (three.js from a CDN, no external models); median 40,747 chars
Teacher	undisclosed teacher model (answers); a second undisclosed model writes the traces

• 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.
• topic is always spatial; language is English; verification is confirm (no planted mistakes in this set).
• difficulty is the requested structural complexity of the object (simple → easy, moderate → medium, intricate → hard).

Difficulty	Rows	Median reasoning tokens (cl100k_base)
easy	251	4,735
medium	412	4,840
hard	308	4,806

How it was built

1. 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.
2. 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/…).
3. Runtime check. Every answer is loaded in headless Chromium (software WebGL), every button is clicked so all three animations run, and the row is dropped on any page error, console error, NaN geometry warning, missing canvas, or an error message shown on the page. About 18% of teacher answers failed this (syntax errors such as -x**2, duplicate let, mergeGeometries with incompatible attributes) and are not in the set.
4. Reasoning. The teacher's own reasoning is not available, so the trace is reconstructed from the finished, working code as an engineer's notes. Shape: silhouette, decomposition into primitives with dimensions, attachments and pivots, the three animations, materials and performance. Code: structure and build order, each helper's inputs (2D outline vs 3D points, radians, indexed vs non-indexed), which variables each loop defines and that nothing is read before its declaration, how merged geometries are made compatible, mode switching, the frame loop. Then 验： with two real checks: spatial arithmetic (stacking contact, camera framing, pivot placement) and a code check that walks one risky spot by name (scope, helper input shape, merge attributes, operator precedence, declaration order).
5. Gates. A trace is rejected unless it is mostly Chinese (measured per word, so code identifiers count once), contains no pasted code, has a 验： that shows a computation and names an identifier that exists in the code, contains no refusal or meta text (v1.7 traces often opened with "I can give a design summary, not my internal reasoning"), and at least 80% of the measurements it quotes before 验： appear in the HTML.

Train on the GPT turn (reasoning + answer); mask system and human.

from datasets import load_dataset
ds = load_dataset("sriq-ai/exp-n950-spatial", split="train")