General Instruction-Following Dataset (750 pairs)

750 original instruction/response pairs (instructions.jsonl) for instruction fine-tuning and evaluation, spanning 7 task types at roughly 107 records each: summarization, rewriting (tone/style transfer), extraction (pulling structured facts from prose), classification (assigning a label from an instruction-provided set), reasoning (multi-step word problems and logic puzzles with shown work), transformation (format conversion — prose to table, list to paragraph, and similar), and structured output (valid JSON/YAML/Markdown conforming to a given spec). Subject matter is deliberately varied across dozens of everyday domains (business, science, history, technology, cooking, travel, health, and more) so the set doesn't overfit to one topic or phrasing pattern. Each record is a flat object with an instruction, a response, and a task_type label; response length varies naturally by task rather than following a uniform template. Contains no real personal data, credentials, or payment information — all names, companies, and figures are fictional or illustrative. Intended for supervised fine-tuning, task-balanced sampling, held-out per-category evaluation, or as a format/constraint-following regression check after a fine-tune.