LuminaSFT: Generating Synthetic Fine-Tuning Data for Small Language Models

2026-02-24 · Source: AMD ROCm Blogs · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, long

Summary

LuminaSFT is a new supervised fine-tuning (SFT) dataset designed to enhance small language models (SLMs), presented on February 24, 2026. The research explores two primary methods: regenerating existing SFT data using a stronger teacher model like DeepSeek-V3 and generating new task-specific data. Experiments show that data regeneration can improve average performance by up to ~4% across tasks like MMLU and GSM8k, with a maximum improvement of ~7% on GSM8k for Instella-3B-base. Task-specific data generation yields more substantial gains, boosting performance by up to ~41% on reading comprehension tasks like DROP for Llama-1B. The study also demonstrates that useful data can be generated from scratch for educational QA tasks using detailed prompts and multi-step pipelines, achieving up to ~2.4% average gain.

Key takeaway

For NLP engineers developing or deploying small language models, consider integrating LuminaSFT's methodologies to boost performance. Regenerating existing SFT data with a powerful teacher model can offer moderate improvements, but focusing on generating task-specific datasets will yield substantially higher gains, particularly for specialized applications like reading comprehension. Even without initial seed data, structured prompting and multi-step pipelines can effectively bootstrap high-quality training data, optimizing your SLM's capabilities for targeted tasks.

Key insights

Synthetic data generation and regeneration significantly improve small language model performance, especially with task-specific approaches.

Principles

• Teacher model strength impacts data regeneration efficacy.
• Task-specific data yields greater performance gains.
• Detailed prompts enable data generation without seed data.

Method

The method involves regenerating SFT data using a stronger teacher model (e.g., DeepSeek-V3) or generating task-specific data from seed datasets or detailed prompts, followed by fine-tuning SLMs.

In practice

• Use DeepSeek-V3 for SFT data regeneration.
• Prioritize task-specific data generation for large gains.
• Employ multi-step generation for educational QA tasks.

Topics

• Small Language Models
• Supervised Fine-Tuning
• Synthetic Data Generation
• Teacher-Student Learning
• AMD Instinct GPUs

Code references

• AMD-AGI/Instella
• allenai/olmes
• hiyouga/LlamaFactory

Best for: NLP Engineer, AI Scientist, Research Scientist, AI Engineer, Machine Learning Engineer, AI Researcher

Related on AIssential

• Learning What to Learn: Stage-Specific Data Sets for SFT-then-RL in Small Language Model Reasoning — Data strategy for SFT-then-RL pipelines should align with each stage's distinct role in skill acquisition and consolidation.
• Synthetic Data for any Differentiable Target — DPG optimizes synthetic data generators using higher-order gradients to precisely control target model behavior via SFT.
• Half hour of labeling power: Can we beat GPT? — LLMs can create high-quality datasets to train smaller, more accurate, private models for specific business needs.
• How Can We Synthesize High-Quality Pretraining Data? A Systematic Study of Prompt Design, Generator Model, and Source Data — Structured pedagogical formats and diverse outputs are key to high-quality synthetic pretraining data, not larger generator models.
• Teaching Language Models to Check Grounded Claim Factuality with Human Test-Taking Strategies — Formulating factuality checking as a true/false reading comprehension task with test-taking strategies improves LLM efficiency and accuracy.
• Children's English Reading Story Generation via Supervised Fine-Tuning of Compact LLMs with Controllable Difficulty and Safety — Fine-tuning compact 8B LLMs can generate children's stories with controllable difficulty and safety, outperforming larger models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AMD ROCm Blogs.