0.8.0

IMPORTANT CHANGE!
The API now runs on port 8338 by default to avoid clashing with other software that was running on port 8000. This new build of the App will look for the server to be on 8338 from now on.

What's Changed

• One of our biggest releases in terms of changes
• On the training page, show the final status of the train (so a train that fails but is complete will show the failure)
• Connect to API faster on Local Connection
• Change default API port to 8338
• Better download progress bars on model download
• Improve Computer Information page
• Show a mini NVIDIA logo if you have an NVIDIA device
• New Logo
• We have a new Log Viewer that is at the bottom of the page -- it streams logs for important tasks
• Fixed the training status text when a failure occurs from success to failure by @sanjaycal in #210
• Update the error message and automatically download a dataset when a part of a recipe is not found by @sanjaycal in #208

New Contributors

• @sanjaycal made their first contribution in #210

Full Changelog: v0.7.5...v0.8.0

0.7.5

Installer improvements:

• Console now works on Windows!
• Fixes for a number of hiccups on Windows
• Improved logging during install
• Fix for slowness/freeze when trying to connect to API

Other improvements:

• Don’t edit stored HuggingFace token if HuggingFace login is already setup and working
• Numerous library and dependency updates
• Updated App icon!

0.7.2

v0.7.2

Bump version to 0.7.2

0.7.1

v0.7.1

Bump version to 0.7.1

0.6.2

• Recipes: alpha release of recipes allows you to import/export training templates via yaml file
• Improved training data visibility: added fields to preview datasets and samples of rendered templated output
• Support repeat for batch completions

0.6.1

v0.6.1

Bump version to 0.6.1

0.6.0

• Simpler installer: Performance improvements and app no longer pauses for user input between steps
• Tool Use: Early preview of Tool Use / Function Calling in chat with demo weather and calculator tools
• Batch Prompting: support for executing batches of prompts simultaneously
• Tokenization Visualizer: Tokenize screen to analyze how a model will tokenize a string
• Interact Dropdown: Consolidated interaction screens in one screen (incl. "Query Docs", "Embeddings" and many new screens)
• Better Streaming: Fixes for streaming on "Completions" and "Prompt Templates" pages
• Export to Hugging Face

0.5.0

v0.5.0

Bump version to 0.5.0

0.4.0

What's New

• Addition of several new training plugins that support a variety of methods including RLHF, DPO, ORPO, SIMPO and SFT.
• Initial support for multimodal models
• Improvements to training interface, including ability to stop a training job
• Plugins now have readme pages that are displayed in the app

0.3.9

• Added Qwen 2 models to the gallery
• Added a number of popular datasets to the gallery
• Update training templates to use Jinja => requires update to existing training templates
• You can now edit training templates
• First UI updates to reorganize training template creation
• Added a Gated badge to gated models and an explanation and link to any agreements required to use gated models
• Generated models have a separate page in the model zoo
• Bug fixes: issues starting up Mistral v0.3 models, downloading models can cause a database lock