Bring data
Browse Hugging Face datasets, import local text or JSONL, mix sources, and track progress on big downloads.
SwiftUI / Apple MLX / real training loop
Build a small GPT. Train it. Talk to it.
LabLLM is a native macOS GUI for training small Transformer models from scratch, fine-tuning behavior, inspecting checkpoints, and chatting with the model you trained.
From scratch GPT-style training
Orange val loss beside blue train loss
Local chat with saved checkpoints
tokens/sec
0
checkpoint
step 1,240
mode
fine-tune
The whole experiment loop
Move from data to a talking checkpoint without rebuilding your workflow.
Design the model
Choose a GPT-style preset, tune architecture details, estimate memory, and let Simple mode handle the tokenizer.
Train and watch
Follow train loss, validation loss, throughput, checkpoints, ETA, and live samples as the model changes.
Sample, chat, continue
Generate inline, chat with checkpoints, continue useful runs, export cards, and keep experiments organized.
Product, not promise
The dashboard is the lab notebook.
Training should not feel like staring at a terminal and hoping the curve means something. LabLLM keeps metrics, validation, samples, checkpoints, and model management close together.
- Blue train loss and orange validation loss are separate curves.
- Samples appear as a timeline during training and fine-tuning.
- Checkpoints can be loaded, renamed, continued, compared, and exported.
Tiny live-feeling previews
A few moving pieces from the workbench.
Training metrics
step 0
train loss
validation loss
Inline sampling
temperature 0.8
Prompt
The model opened its eyes and noticed
Data browser
importable sources
Checkpoint timeline
best run marked
240
520
880
1240
selected checkpoint
story-gpt-sft-0880
val loss improved, sample quality up, ready to continue or chat.
Model chat
streaming
Explain why validation loss matters.
Hugging Face import
chunked
Tiny Stories 20%
validating filesjsonl/parquet
preview rows2,000
Screens from the app
Real tools for real experiments.
Bring useful data
Search, inspect dataset cards, import actual trainable files, and mix sources with row limits or percentages.
Keep the good runs
Load, continue, rename, quantize, and compare checkpoints without losing the thread of the experiment.
Roadmap tracker
Built, being hardened, and planned.
LabLLM has a lot of surface area, so the roadmap separates working beta features from active correctness work and bigger future ideas.
Open ROADMAP.mdPull requests welcome
Help make small-model training easier to trust.
The highest-impact work is focused: tests, ML correctness, dataset edge cases, SwiftUI polish, docs, screenshots, recipes, and better explanations for errors.
Quick answers
Useful before you download.
Does it train real models?
Yes. LabLLM builds GPT-style decoder models with Apple MLX and saves checkpoints you can load back into the app.
Can the models talk naturally?
Yes, with good instruction or conversation data, you can fine-tune behavior and chat with the result inside LabLLM.
What Mac do I need?
macOS 14 or newer on Apple Silicon. More memory gives you more room for larger experiments.
Where should contributors start?
Core correctness tests, dataset import reliability, tutorial polish, and focused SwiftUI fixes are the best entry points.
Open source beta