Supervised Fine-Tuning Dataset For Language Model Training

SFT-Dataset is a collection for supervised fine-tuning of base language models like Qwen/Qwen3-4B-Base. The dataset is tagged for text generation, reasoning, math, science, and code tasks. It is authored by 96kevinli29 and was last updated in March 2026.

Use Cases

• Fine-tune a base language model on text generation tasks using the SFT-tagged instruction-response pairs.
• Train a model for multi-domain reasoning by leveraging the dataset's math, science, and code modality tags.
• Use the supervised fine-tuning mixture to create a policy model as a foundation for subsequent reinforcement learning.

Strengths

• Dataset is explicitly designed for supervised fine-tuning, as indicated by its 'Sft' and 'Supervised Fine Tuning' tags.
• Covers multiple analytical domains including text generation, reasoning, math, science, and code.
• Last update recorded as March 24, 2026, indicating recent maintenance.

Limitations

• Core metadata such as row count, column structure, and sample data is unavailable for assessing scale and content.
• Specific license terms are unknown, which could restrict commercial or research use.
• The description references upstream mixes and target counts but does not provide the actual composition or quality metrics.

Provenance

Source

Hugging Face, uploaded by author 96kevinli29.

Collection Method

Curated mixture for supervised fine-tuning; specific gathering method is not detailed.

Freshness

Last updated on 2026-03-24.

Geography

Tagged with region 'us', but specific spatial coverage is not detailed.