Pondhouse Data Blog

What Vertex AI Model Garden is and how to use it: Claude 3.5 Sonnet and Llama 3.1 as pay-per-use API services, CLIP on a VM, and Hugging Face models.

Fine-tune GPT-4o mini in the OpenAI dashboard: the JSONL training format, how many examples you need, job settings, costs and calling the model from Python.

Set up the Neo4j LLM Knowledge Graph Builder on AuraDB or Docker, turn a web page into a knowledge graph and query it with GraphRAG in the built-in chat.

Finetune Llama 3 8B on a single GPU with QLoRA, bitsandbytes and TRL: dataset formats, LoRA and training parameters, full Python script and inference.

Call OpenAI from PostgreSQL with the pgai extension: create embeddings, summaries and content moderation in plain SQL, plus triggers that embed new rows.

How Kolmogorov-Arnold Networks (KANs) replace MLP weights with learnable splines, what the paper shows on accuracy, and why MLPs aren't done yet.

Llama Guard 2 is Meta's 8B model for checking LLM prompts and responses. The main attack vectors, its 11 risk categories and how to run it.

Let GPT-4 write python-pptx code that builds a PowerPoint deck from your company documents via RAG, using your own corporate template.

What Vertex AI Agent Builder is and what Google calls it today, with a no-code example: index Slack messages and build a Gemini search app.

How Microsoft's BitNet trains LLMs with 1-bit and ternary weights, why it beats post-training quantization, and what it saves in memory and energy.