SDG_Hub: An Open-Source Toolkit for Synthetic Data Generation & LLM Customization
summary
SDG Hub is an open-source toolkit that composes prompts, blocks, and flows into reusable, YAML-or-Python pipelines that scale automatically and wrap LLM calls or pure Python transforms. It claims knowledge tuning via three augmentations (thematic summary, key-fact lists, and connected facts) to turn private docs into instruction-tuning data, demonstrates a doc-grounded skill and a reasoning customization, shows Llama 3 18B improving precision after fine-tuning, and now supports GPOSS 12B as teacher plus a GUI for visual pipeline editing.
AI-written summary · about
chapters (17)
- 0:00SDG Hub introduction
- 1:05Why synthetic data for LLMs
- 3:52Open-source model advantages
- 5:17Synthetic data challenges
- 7:46SDG Hub modular blocks
- 10:34Quick start guide
- 11:39Customizing SDG Hub
- 12:46Domain-specific LLM customization
- 13:28Domain knowledge importance
- 17:22Knowledge tuning basics
- 19:51Raw document training limits
- 21:14Document augmentation methods
- 23:18Teacher-student model roles
- 25:06Fine-tuned model demo
- 26:53Reasoning in fine-tuned models
- 28:19SDG Hub technical features
- 29:23Q&A
related talks
28mA community based open source LLM and what it can do for your business
46mBeyond Basic Prompting: Supercharging Open Source LLMs with LMQL's Structured Generation
49mNext-Gen Documentation with AI
21mThe Llama Toolbox: A Collection of Practical Tools and Recipes
40mInstructLab: Applying Open Source Methods to Building and Trai...
15mAn LLM Toolchain for Open Data and NGSI-LD
28mThe Data That Shapes Foundational LLMs
30mApplying Open Source Methods to Building and Training Large Langu...