Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
QLoRA Fine-tuning AI Vtuber Identity
Step‑by‑step guide to building a real‑time AI VTuber using LLaMA‑3.2 3B fine‑tuned with QLoRA, covering dataset prep, prompt design, LoRA layers, quantization, and RTX‑3060 optimization.
Can Identity be fine-tuned with QLoRA for an AI Vtuber?
In this technical demonstration, I’ll walk you through the step-by-step creation a real-time interactive VTuber powered by a base LLM (LLaMA 3.2) 3B*, fine-tuned with QLoRA and the Unsloth framework. You’ll see firsthand how we went beyond typical “instruction-following” models to craft a uniquely creative and reflexive virtual personality.
I’ll share:
How we efficiently applied LoRA layers for lightweight, GPU-friendly fine-tuning.
Deep insights into dataset preparation, custom prompt engineering, and careful handling of conversational structure to create a convincingly human-like personality.
Technical challenges overcome, including managing VRAM constraints (on an RTX 3060) through quantization, CPU offloading, and careful gradient management.
Demonstrates LLaMA 3.2 fine-tuned for real-time, reflexive AI VTuber interaction.
Related projects
from LLMs to reasoning models
Quito
This talk covers implementing reasoning models by scaling inference-time computation on open source LLMs using techniques like Monte…
Reinforcement Learning in Action: Building a Q-Learning System for Real-World Inventory Optimization
Quito
Live code walkthrough of Q‑Learning applied to inventory optimization, covering reward design, Q‑table mechanics, state transitions, and practical…
Building a Unified AI Interface: Live Demo of Dolphin MCP's Cross-Provider Tool Orchestration
Quito
This talk demonstrates building a unified AI interface using Dolphin MCP to orchestrate tools across multiple LLM providers…
SomosNPL
Quito
The talk covers efforts to advance Spanish natural language processing by creating open resources, highlighting SomosNPL's role in…
AnythingLLM: la plataforma integral para ejecutar modelos LLM locales y agentes de IA
Manizales
Demo of AnythingLLM: installing a local AI app, loading GGUF models, using RAG with PDFs/Word/CSV, and connecting OpenAI,…
Auto-create reliable LLM evals
New York City
Learn how to build reliable LLM-based evaluations using only about twenty human annotations, achieving scalable, human‑aligned assessments that…