v0.7.3

Hotfix for MLX issue between latest logprob changes and latest version of MLX. Also includes a number of logprob updates.

v0.7.2

• aligned job status to always use RUNNING instead of mix of RUNNING and IN_PROGRESS
• recipes import returns a list of assets and whether they need to be downloaded or not
• updated model architecture to MLX and LlamaFactory trainers
• jobs debug information for settings page

v0.7.1

Hotfix to match updated function signature for generate_step in MLX server.

v0.7.0

• Llama.cpp/GGUF Fixes: Fixed performance issues for systems with GPU, GGUF models correctly show as downloaded, issues with inference for models with stop_string fixed
• Model gallery: Added “new” badge and tags, older models archived by default
• MLX: Libraries updated. Qwen 2.5 support, more LoRA params supported, validation loss added to tensor board
• Recipes: Gallery added to import screen
• Datasets: support for showing specific splits in preview

v0.6.8

Initial support for importing and exporting recipes.

v0.6.7

Bumped Llamafile version to latest: v0.8.16

v0.6.6

• Llamafile Export: You can now export any GGUF model to a self-contained self-executing Llamafile
• Several Import/Export bug fixes: Mostly for GGUF, plus better error handling and more verbose debug output

v0.6.5

Patch release to support additional validation split names.

v0.6.4

Patch release to add automatic creation of a validation split from training data in MLX trainer plugin.

v0.6.3

If you are training on Apple devices, this update contains important fixes.

• MLX Trainer:
  • output and tensorboards work again
  • Fixed bugs with validation split not being read properly and added a check for required splits before starting training.
• Report empty Datasets

Full Changelog: v0.6.2...v0.6.3