A curated, medium-scale mixture designed to push a base model toward two things at once: stronger step-by-step reasoning (math, science, code) and reliable instruction following (format, language, and task constraints).
Quantities are chosen to stay trainable on modest GPU budgets while keeping signal density high—useful as a standalone SFT stage or as a clean warm start before reinforcement learning.
Evidence: benchmarks on a model trained on this mixture… See the full description on the dataset page: https://huggingface.co/datasets/SeaFill2025/SFT-Dataset.
