SFT Dataset: A Mixture for Instruction Following and Step-by-Step Reasoning

SFT-Dataset is a curated, medium-scale mixture designed to push a base model toward stronger step-by-step reasoning and reliable instruction following. The dataset was created by SeaFill2025 and was last updated on Hugging Face in April 2026. Quantities are chosen to be trainable on modest GPU budgets while keeping signal density high.

Use Cases

• Fine-tuning models for step-by-step reasoning based on described math and science content
• Improving model adherence to task and format constraints based on described instruction-following goals
• Warming up a model before reinforcement learning based on the described use as a clean starting point
• Training models for code generation based on the code-related content mentioned in the platform tags

Strengths

• Designed as a curated mixture for two specific training objectives: reasoning and instruction-following
• Quantities are chosen to be trainable on modest GPU budgets
• Last updated on 2026-04-03, suggesting recent maintenance

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count, file formats, and exact size are unknown, which may limit suitability assessment
• License information is unknown, which could restrict usage

Provenance

Source

SeaFill2025 on Hugging Face

Collection Method

Curated mixture; specific gathering method is not detailed.

Time Range

null

Freshness

Last updated 2026-04-03 12:41:09

Geography

null