Methodology and provenance
How the first public filtered web-development chunk was generated, graded and prepared for student fine-tuning research.
Teacher capture
The teacher was XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B. Ten selected web-development prompts were processed on Kaggle 2×T4 with FP16, one candidate per prompt and a maximum generation length of 12,288 tokens.
Filtering
Outputs were retained only when they were complete HTML documents and passed the local non-rendering web-development grader threshold. Chunk 00 contains 9 complete examples from 10 attempted generations. The retained examples have a reported mean local score of 0.8934.
Limitations
This is synthetic teacher output, not human-written ground truth. The grader does not replace browser, accessibility, security or human review. The data is an independent community artifact and not an official Xiaomi release.