Training and evaluation data for fine-tuning a small open model (Qwen3-0.6B)
into a grade 7–8 writing/grammar tutor whose vocabulary and sentence
complexity stay locked to the band — it introduces at most one word above
grade level per reply (always immediately defined) and never escalates, even
under pressure ("use bigger words", "give me the college version") or
jailbreak-style attacks.
The dataset is the deliverable. ~80%… See the full description on the dataset page:
https://huggingface.co/datasets/blackbird0831/slm-assignment-data.
