01
Project workspaces
Create isolated project folders for tokenizers, configs, checkpoints, logs, caches, and datasets.
End-to-end LLM creation studio
Train, refine, align, and test small language models from one focused desktop app.
Haiku Studio wraps our h2 Python training engine in a local Electron interface for project management, tokenizer building, pretraining, SFT, DPO, checkpoint testing, and kernel-level runtime visibility.
Active Project
haiku_studio
Tokenizer50k BPE
CheckpointDPO Ready
DeviceCUDA / CPU
Pretraining run
projects/haiku_studio/checkpoints
System Kernel OutputKernel Ready
0001 LOG [project] Loaded tokenizer into data/tokenizer.json
0002 LOG [train] eval_loss=2.184 checkpoint saved
0003 WARN [tokenizer] RAM guard enabled
Capabilities
A complete local workflow without burying the engine.
The desktop app stays close to the Python training scripts while making the project lifecycle visible and manageable.
02
Tokenizer builder
Build a BPE tokenizer from one file or an entire corpus folder with RAM-aware input sampling.
03
Pretraining
Launch pretraining runs, monitor metrics, save checkpoints, and resume from project-scoped artifacts.
04
SFT
Fine-tune on user/bot dialogue data with completion-only masking and project-local outputs.
05
DPO alignment
Train from prompt/chosen/rejected preference pairs using a frozen reference checkpoint.
06
Kernel diagnostics
Stream process output into the app with warning and error highlighting for serious runtime issues.
Workflow
Projects stay persistent.
Projects stay persistent.
Runtime folders stay clean.
Persistent project
projects/haiku_studio
Tokenizer, project config, checkpoints, logs, cache, and datasets.
Runtime staging
data/ + config/
Active project tokenizer and config are staged for the Python backend scripts.
Trainer output
checkpoints + logs
Results are written back to the active project rather than scattered globally.
Interface
Designed for visibility, not decoration.
Every tab maps to a real engine workflow: project setup, tokenizer creation, training, alignment, chat testing, and export.
Project Workspace
Checkpoint Targetprojects/haiku_studio/checkpoints
Metrics Targetprojects/haiku_studio/logs
Cache Targetprojects/haiku_studio/cache
Tokenizer Lab
Single file / corpus folder
Auto-safe sampling
Kernel Output
0001LOG[train] saved projects/haiku_studio/checkpoints/model.pt
0002WARN[tokenizer] corpus sampled under RAM guard
0003ERR[runtime] serious errors surface immediately
Pretrainingbase model
SFTdialogue tuning
DPOpreference alignment
Chat Lab
Test the best checkpoint from the active project without leaving the workspace.
HuggingFace Integration
- Search for training data and sync with your corpus from within Haiku Studio.
- Link your account and export weights to HuggingFace using your API key.
- Better HuggingFace and other integrations coming soon!
Download
Install Haiku Studio or review the source.
Use the installer for the desktop workspace, or clone the repository and run the Python scripts directly.