Dashboard
Good afternoon, Jon
Tokens / sec
2,412
Active runs
3
Saved
$1,240
Recent activity
A friendly, technical brand book inspired by Unsloth Studio. Mint green energy on clean off-white, anchored by bold black accents. Every token, component, and pattern you need to build a fast, modern AI/ML product.
Logomark
Mint disk + wordmark. Always pair with at least 8px of clearspace.
Voice
Design principles
Mint scale (primary)
Neutrals
Semantic
Category accents
Weights
Spacing
Radius
Shadow
Token must start with hf_.
Precision
Advanced
Send telemetry
Anonymous training metrics.
Total optimizer steps. Use 0 to run by epochs.
Closed (trigger)
Open
KPI cards
Tokens / sec
VRAM used
Active runs
2 LoRA · 1 QLoRA
Cost saved
vs FA2 baseline · this month
Feature cards
Import from Hugging Face, CSV, or JSON. Auto-format prompts and labels.
Sane defaults for LoRA, QLoRA, full FT. Override anything from the sidebar.
Loss, throughput, VRAM. Real-time charts that don't lag.
Complex card · training run
Started 14m ago · 7.5k / 12k steps · ETA 9m
Loss
1.04
Tokens / sec
2,412
VRAM
14.2 GB
Learning rate
5e-5
Pricing cards
For tinkering and learning.
For serious fine-tuners.
For production training at scale.
Profile card
Staff ML Engineer · sloth-team
Data table
| Run | Model | Status ↓ | Loss | Tok/s | Owner | Updated | ||
|---|---|---|---|---|---|---|---|---|
| #482 | qwen3.5-4b | RUNNING | 1.04 | 2,412 | JR Jon R. |
14m ago | ⋯ | |
| #481 | llama-3-8b | DONE | 0.87 | 1,890 | DK Daniel K. |
2h ago | ⋯ | |
| #480 | gemma-4-2b | PAUSED | 1.22 | 3,104 | AL Alex L. |
5h ago | ⋯ | |
| #479 | mistral-7b | FAILED | — | — | MP Maya P. |
1d ago | ⋯ | |
| #478 | qwen3-1.5b | DRAFT | — | — | JR Jon R. |
2d ago | ⋯ |
Showing 1–5 of 42
Compact / striped
| Hyperparameter | Default | Current | Range |
|---|---|---|---|
| learning_rate | 2e-4 | 5e-5 | 1e-6 → 1e-3 |
| batch_size | 8 | 16 | 1 → 128 |
| max_steps | 100 | 200 | 0 → 100000 |
| lora_rank | 16 | 32 | 1 → 256 |
| warmup_ratio | 0.03 | 0.05 | 0 → 0.5 |
Training complete.
qwen3.5-4b finished after 14m 28s. Final loss 0.87.
VRAM approaching limit.
14.2 / 16 GB used. Consider enabling gradient checkpointing.
Run #479 failed.
CUDA out of memory at step 312. View logs for stack trace.
New model available.
qwen3.6 dropped today. Drop-in compatible with your current config.
Unlock multi-GPU training and 2.5× speedups.
Underline
Pill (segmented)
Chip filter
qwen3.5-4b on alpaca-cleaned. Estimated 14 minutes on your current GPU.
Run started
#482 · qwen3.5-4b
Checkpoint saved
step-7000.safetensors · 2.1 GB
Linear
Circular
Training
7,520 / 12,000 steps
ETA 9m 22s
Skeleton
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = "unsloth/qwen3.5-4b",
max_seq_length = 2048,
load_in_4bit = True,
)
trainer.train() # 2× faster, 50% less VRAM$ unsloth train --model qwen3.5-4b ✓ Loaded base model (4.0B params) ✓ Mounted dataset alpaca-cleaned (52,002 rows) ⚡ Starting LoRA fine-tune... step loss tok/s vram ──── ──── ───── ──── 0500 1.42 2,310 12.1G 1000 1.21 2,398 13.8G 7520 1.04 2,412 14.2G ← live
Throughput by model
tokens / sec
Training loss
run #482
VRAM allocation
Kick off your first training run by choosing a model and a dataset. We'll keep you posted.
Dashboard
Tokens / sec
2,412
Active runs
3
Saved
$1,240
Recent activity
Contrast ratios against white background. WCAG AA target ≥ 4.5 for body text, ≥ 3 for UI / large text.
| Foreground | Background | Contrast | Use |
|---|---|---|---|
| Carbon #0A0A0A | White | 19.8 ✓ | Body, headings |
| Slate Ink #475569 | White | 8.0 ✓ | Secondary text |
| Deep Mint #0FA968 | White | 4.7 ✓ | Mint links, labels |
| Sloth Mint #22D88F | White | 2.3 △ | UI only, ≥ 18pt large text on white |
| Carbon #0A0A0A | Sloth Mint | 8.6 ✓ | Primary CTA (carbon-on-mint) |