How h2oai/h2o-llmstudio works
H2O LLM Studio is a GUI and CLI framework for fine-tuning LLMs. Users configure a project and dataset, create and launch an experiment, monitor and compare results, then chat with or export a trained model. The graph captures the GUI workflow and training/evaluation layers; the README also documents LoRA, DPO, multiple problem types, W&B tracking, and Hugging Face export. Some subsystem links use README-level evidence because their implementations were not sampled.
Main entry points: train.py Training orchestrator · handlers.py UI handlers · experiment.py Experiment UI · dataset.py Dataset UI
Components 25
GUI and projects
- Wave app
llm_studio/app.py - UI handlers
llm_studio/app_utils/handlers.py - Home and projects
llm_studio/app_utils/sections/home.py - Dataset UI
llm_studio/app_utils/sections/dataset.py - Experiment UI
llm_studio/app_utils/sections/experiment.py - Model chat
llm_studio/app_utils/sections/chat.py - User settings
llm_studio/app_utils/sections/settings.py
Data and configuration
- App configuration
llm_studio/app_utils/config.py - Experiment database
llm_studio/app_utils/db.py - Dataset preparation
llm_studio/src/utils/data_utils.py - Conversation chains
llm_studio/src/datasets/conversation_chain_handler.py - Tokenization and text
llm_studio/src/datasets/text_utils.py - Problem configurations
llm_studio/python_configs - Config conversion
llm_studio/src/utils/config_utils.py
Training and evaluation
- Training orchestrator
llm_studio/train.py - Modeling utilities
llm_studio/src/utils/modeling_utils.py - Task models
llm_studio/src/models - Task losses
llm_studio/src/losses - Evaluation metrics
llm_studio/src/metrics
Results and integrations
- Training plots
llm_studio/src/plots - Logging and charts
llm_studio/src/utils/logging_utils.py - Model export
llm_studio/src/utils/export_utils.py - Hugging Face Hub
- Weights & Biases
Other components
- User
Connections 27
- User → Wave app: uses
README.md - Wave app → UI handlers: dispatches
llm_studio/app.py - UI handlers → Home and projects: routes to
llm_studio/app_utils/handlers.py - UI handlers → Dataset UI: routes to
llm_studio/app_utils/handlers.py - UI handlers → Experiment UI: routes to
llm_studio/app_utils/handlers.py - UI handlers → Model chat: routes to
llm_studio/app_utils/handlers.py - UI handlers → User settings: routes to
llm_studio/app_utils/handlers.py - Home and projects → App configuration: reads
- Dataset UI → Dataset preparation: prepares data
llm_studio/app_utils/sections/dataset.py - Dataset UI → Conversation chains: uses
llm_studio/app_utils/sections/dataset.py - Conversation chains → Tokenization and text: calls
llm_studio/src/datasets/conversation_chain_handler.py - Experiment UI → Problem configurations: configures
llm_studio/app_utils/sections/experiment.py - Experiment UI → Config conversion: uses
llm_studio/app_utils/sections/experiment.py - Experiment UI → Training orchestrator: starts
- Training orchestrator → Config conversion: loads config
llm_studio/train.py - Training orchestrator → Dataset preparation: loads data
llm_studio/train.py - Training orchestrator → Modeling utilities: uses
llm_studio/train.py - Modeling utilities → Task models: supports
llm_studio/src/models/text_dpo_modeling_model.py - Modeling utilities → Task losses: supports
- Training orchestrator → Evaluation metrics: evaluates with
- Training orchestrator → Logging and charts: logs
llm_studio/train.py - Training orchestrator → Training plots: produces plots
- Experiment UI → Experiment database: reads and writes
- Model chat → Task models: chats with
- Experiment UI → Model export: exports model
llm_studio/app_utils/sections/experiment.py - Model export → Hugging Face Hub: publishes to
README.md - Training orchestrator → Weights & Biases: tracks runs
README.md