ToolBunny Logo
ToolBunny
Back to Homepage

Hyperparameter Visualizer

Map context weights interactively to conceptualize fine-tuning parameter scales.

Target Foundation Model Architecture:
Interactive Scale Adjustment
Learning Rate ($\eta$)1.00e-4 (10^-4)
1e-6 (Slow)1e-4 (Default LoRA)1e-2 (Exploding)
LoRA Rank ($r$)16
Adapts 16.8M trainable params (0.210%)
LoRA Alpha ($\alpha$)32
Scaling Ratio $\alpha / r$: 2.00x
Per-Device Batch Size4
Effective Batch Size: 16 (4 x 4)
Training Epochs3
Total full passes over training dataset
Inference Temperature ($T$)0.7
0.1 (Deterministic)0.7 (Creative)2.0 (Random)
Simulated Training Loss Curveoptimal Convergence
Estimated VRAM Memory Footprint (Llama-3 8B)
QLoRA 4-bit
~7.6 GB
1x RTX 4090 (24GB)
LoRA 16-bit
~22.1 GB
1x A100 (40GB)
Full Fine-Tune
~144.0 GB
Multi-GPU Cluster
Softmax Token Probability Shift (Temperature = 0.7)
Token A78.9%
Token B12.3%
Token C7.0%
Token D1.1%
Token E0.5%
Token F0.2%
Export Configuration Code
from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "llama-3-8b",
    max_seq_length = 2048,
    load_in_4bit = True,
)

model = FastLanguageModel.get_peft_model(
    model,
    r = 16,
    lora_alpha = 32,
    lora_dropout = 0,
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
)

# Training Arguments
learning_rate = 1.00e-4
effective_batch_size = 16 (4 x 4)
epochs = 3

About This Tool

The Hyperparameter Visualizer is an interactive simulation suite built for machine learning engineers, AI researchers, and prompt engineers who fine-tune or sample Large Language Models (LLMs).

Understanding how fine-tuning hyperparameters interact — such as how LoRA Rank ($r$) and Alpha ($\alpha$) dictate trainable parameter counts, how exponential Learning Rate ($\eta$) scaling influences training convergence, and how Softmax Temperature ($T$) flattens vocabulary token probability distributions — is critical to avoiding gradient explosion, underfitting, or out-of-memory (OOM) GPU crashes. This tool visualizes parameter scales, simulates loss curves, computes VRAM memory footprints, and exports Unsloth and HuggingFace TRL code configurations.

How to Use

  1. Select Target Foundation Model: Choose between Llama-3 8B, Qwen 2.5 14B, DeepSeek-R1 32B, or Llama-3 70B parameter presets.
  2. Adjust Fine-Tuning Sliders: Move the exponential Learning Rate slider, set LoRA Rank ($r$) and Alpha ($\alpha$), and configure batch size and epochs.
  3. Observe Real-Time Visualizations: Watch the simulated training loss curve update dynamically (indicating optimal, underfitting, or divergent behavior), inspect trainable parameter percentages, and review QLoRA vs Full Fine-Tune VRAM estimates.
  4. Explore Temperature Probability Shift: Drag the Inference Temperature slider to visually see how token softmax probabilities flatten or sharpen.
  5. Export Code: Copy ready-to-paste Python code for Unsloth, HuggingFace TRL `SFTTrainer`, or OpenAI/vLLM sampling JSON payloads.

Common Use Cases & FAQ