Hello Readers,

I'm excited to announce that The Forensic Lunch is back with another episode! This week, we had the privilege of hosting Wyatt Roersma, who shared his insights on training open-source AI models for specialized tasks.

Wyatt has been exploring how to take open-source AI models, like Qwen-2.5, and train them using examples such as YARA rules and targeted prompts to enhance their usefulness for specific applications. In the episode, he walks us through the process step-by-step, empowering you to apply similar techniques to solve your unique challenges.

For instance, I'm currently experimenting with getting AI models to write dfvfs code. While the models are fairly accurate, I believe with a bit of fine-tuning and additional training, they could become even more precise and reliable.

Key Resources from Wyatt's Discussion

Here are some invaluable links to help you dive deeper into the topics discussed in the episode:

• Download Wyatt’s trained models: Hugging Face – vtriple (Wyatt Roersma)
• Rent GPU containers: Vast.ai
• Run open-source models without local hardware: OpenRouter
• Discover and download models: Hugging Face
• Run local models: Ollama
• Train and fine-tune models: LLaMA-Factory GitHub Repository
  • SFT example: Train with LLaMA-3 LoRA SFT
  • DPO example: Train with LLaMA-3 LoRA DPO
  • LoRA example: Train with Qwen2VL LoRA SFT

Watch the Episode

You can catch the full episode below and learn how to start training your own open-source AI models to tackle specialized problems:

Or click the link here:
https://www.youtube.com/live/z6QkYHo97k0