AI Bootcamp: LLM Finetuning & Deployment

On Friday, July 4th, 2025, Float16 in collaboration with the Typhoon SCB 10X team organized the AI Bootcamp: LLM Finetuning & Deployment at DistrictX, FYI Building. This event marked a significant milestone in promoting AI technology development in Thailand. The bootcamp received overwhelming interest and was successfully completed beyond expectations.

Event Overview

The AI Bootcamp was a full-day hands-on training program designed to take participants from basic understanding to practical application of Large Language Models (LLMs) through finetuning and deployment processes using tools and GPUs.

Highlights

🔧 Morning Session: How to Finetune Typhoon Open-Source LLMs

Speaker: Surapon Nonesung, Research Scientist at SCB10x

👉 Key Takeaways from the Typhoon Team: 5 Tips for Fine-Tuning to Get High-Performance Models

1. Spend over 80% of your time on data preparation (because data quality is the foundation of a good model)
2. Create at least 2 evaluation datasets, with at least 1 dataset containing data that has never been trained before
3. Use train set and eval set during Fine-Tuning to check for overfitting
4. Evaluate the model before and after Fine-Tuning to see if the model actually improves
5. Check and adjust the Chat Template used, such as system prompt, instruction format, etc. A good template helps the model answer more accurately and perform much better

🚀 Afternoon Session: How to Deploy Your Finetuned Typhoon Model on Float16 ServerlessGPU

Speaker: Matichon Maneegard, Founder at Float16

👉 Key Takeaways from the Float16 Team: 3 Techniques for Improving LLMs for Real Software Development Use

1. Choose the right LLM file formats for your purpose:
   • .safetensors → For HuggingFace, separate model-weight, tokenizer, architecture files
   • .gguf → For llama-cpp, Ollama, LM-studio - easy to use
   • Choose format based on task:
     • safetensors for fine-tuning
     • gguf for inference
2. Implement OpenAI API Compatibility:
   • Make existing code work without rewriting
   • Change endpoint from OpenAI to use our own model
   • Save costs and have full control
3. Structured Output (Grammar) improves response quality:
   • Use xgrammar, outlines, guidance to control response format
   • JSON mode for accurate function calling
   • Define grammar rules for SQL, option selection, or specific formats

Excellent Response

From post-event evaluations, we received better-than-expected feedback:

💬 Some Impressive Comments

"I was impressed with the overall activity. It was great to learn from people who actually work in the field." - Participant

"The SCB10X and Float16 teams were very attentive in providing knowledge. The management was excellent. Matichon from Float16 organized the Lecture & Workshop format very well." - Participant

"I really enjoyed the event. Especially how smooth and well-streamlined the event was." - Participant

What Participants Received

🎯 Practical Knowledge

• Hands-on experience in fine-tuning and deploying LLMs
• Usage of Typhoon open-source models
• Techniques for improving LLMs for Software Development work

💰 Benefits

• 100% free of charge
• GPU Credits from Float16 worth 1,000 THB
• Digital certificate
• Free lunch

🤝 Network

Met and exchanged experiences with:

• Professors, students, researchers, and Data Scientists
• Engineers and Developers
• Startup founders and entrepreneurs

We thank everyone who participated in the event, including:

• The SCB 10X and NVIDIA teams for their full support
• DistrictX for providing the venue
• All participants who provided feedback and great ideas

Float16 is committed to being part of pushing Thailand to become a regional AI leader. This event is just the beginning of building a strong and sustainable AI practitioners community.

#AIBootcamp #LLMFinetuning #Float16 #SCB10X #NVIDIA #MachineLearning #AIThailand #Typhoon