Open source · Free forever
LLM Fine-Tuning
Visualized in 3D
Interactive explorations of LoRA, QLoRA, DPO, and more. See how each method works, calculate GPU costs, and find the right approach for your project.
Six Methods, One Visual Language
Click any card to explore its 3D visualization
🔥
🔌
📦
📎
🧩
👍
Full Fine-Tuning
Maximum power. Maximum cost.
θ' = θ − α · ∇L(θ)
Params: 100%
VRAM 7B: ~60GB
Save: Full model
LoRA
Low-rank adapters. 1000x smaller checkpoints.
W' = W₀ + BA, B∈ℝ^{d×r}, A∈ℝ^{r×d}
Params: ~0.1-2%
VRAM 7B: ~16GB
Save: 10-50MB
QLoRA
4-bit base + full-precision adapters. 65B on one GPU.
W_NF4 + BA, paged AdamW
Params: ~0.1-2%
VRAM 7B: ~6GB
Save: 10-50MB
Prefix Tuning
Learned soft prompts steering frozen attention.
P = [p₁,...,pₘ] → K, V at each layer
Params: ~0.05-0.1%
VRAM 7B: ~14GB
Save: ~1MB
Adapter Layers
Bottleneck modules between frozen layers.
h ← h + f(h · W_down) · W_up
Params: ~0.5-3%
VRAM 7B: ~18GB
Save: ~30MB
RLHF / DPO
Human preferences sculpt model behavior.
L = −log σ(β·[log π/π_ref(y_w) − log π/π_ref(y_l)])
Params: 100%
VRAM 7B: ~60GB (DPO)
Save: Full model
Built for Engineers
More than just pretty graphics
3D Neural Network Viz
Interactive Three.js scenes showing actual architecture — frozen weights, adapters, gradient flow. Drag to orbit, scroll to zoom.
GPU Playground
Select your hardware, pick a model size, adjust LoRA rank — see real-time VRAM estimates, training cost, and whether your setup fits.
Code Snippets
Copy-paste HuggingFace PEFT code for every method. Real, runnable examples — not pseudocode.