Changelog
Follow up on the latest improvements and updates.
RSS
NEW
- Learning curves are now available in the UI
- Phi-3 now available available for fine-tuning and inference
- Codellama 7b instruct and non-instruct available for fine-tuning and inference
- Dev tier users are now able to deploy up to 8B params (incl. llama-3-8b)
- Ability to list repos in new SDK
- Docs: Added instructions for JSON mode (REST & Python)
IMPROVED
- Improved error messages for incorrectly formatted datasets
- Changed default temperature in prompt UI to 1
FIXED
- Fixed edge-case handling on base model selection in Prompt UI
new
improved
fixed
Release: 4/25/24
NEW
- New fine-tuning stack: Fine-tune on Predibase with up to 10x speedups and improvements in model quality
- Llama-3: All Llama-3 variants are now available for Inference & Fine-Tuning
- Adapters as a first-class citizen: Models have been renamed to Adapters with a new emphasis on adapter-based fine-tuning. Users have the ability to see old model configs as needed.
- New Python SDK:We’re releasing a more consistent and robust SDK for our users to interface with Predibase
IMPROVED
- Docs: Add instruction template section under Fine-tuning > Instruct Formats
- Docs: Added new Quickstart and End-to-End Example
- Training: Allow using model's full context length by default for fine-tuning
- Poll Adapter Status in background while training
FIXED
- Batch size tuning when gradient checkpointing is enabled
- Hide loss column in adapter repos
- Disable mixed-precision training to eliminate OOMs
- Fixed hanging behavior while using generate_stream endpoint
new
improved
fixed
Release: 3/28/24
NEW
- A100's for all: We've heard your feedback and movedalltraining jobs from A10G to A100 GPUs by default!
- Faster Training Times: In addition, we've also modified the default configurations to speed up A100 training by 2x - 5x
- New Tutorial: We've added a new “Build Your Own Lora Land” Tutorial to the homepage and docs
- Mistral-7b-instruct-v0.2 now available: You can now query Mistral-7b-instruct-v0.2 as a Serverless Endpoint as well as use it as for Dedicated Deployments & Fine-Tuning
- Deployments UI: We've added the ability to seamlessly create deployments directly from the UI
- Ability to query Stopped Models: We've changed the "Cancel" operation to "Stop" during training, allowing you to use the model from the latest saved checkpoint.
New Quickstart
Create Deployments via the UI
IMPROVED
Prompt UI Improvements
- We now show all deployments (regardless of the status) in the dropdown
- We show the the status chip next to deployment name in dropdown and in status indicator
- We also provide a "Stop" Button while a response is streaming back to UI
Pricing
- We've updated the serverless pricing bucket (Up to 13B) to include larger models (Up to 21B) at the same price ($0.25 / 1k tokens)
- We've now enabled billing for serverless inference including both the streaming and non-streaming endpoints
FIXED
- Prompt UI: We now prevent streaming of multiple responses at the same time
- Models UI: We fixed the bug where 0 values didn't show in the Learning Curves
- We've fixed the hanging behavior while using the Python SDK in Colab
- We've also improved error messages in the SDK so they're concise and readable
new
improved
fixed
Release: 3/18/24
New
- New Deployments Page in UI for both dedicated deployments and serverless endpoints
- Ability to easily use REST / SDK / CLI code snippets for prompting deployments
- Ability to delete deployments via the UI
- Ability to view live events, real-time deployment statuses and configuration
- Enabled automatic early stopping by default (validation loss over 5 consecutive checkpoints)
Improved
- Docs: Added conversion script for users with OpenAI fine-tuning datasets
- Set "My Profile" as default page on Settings
Fixed
- Improved error handling around creating model repositories with the same name
new
improved
fixed
Gemma Release
New
- Gemma-2B and Gemma-7B added as Serverless Endpoints
- Add Streaming to Prompt UI
Improved
- Show Deployment Status in Prompt UI
- Add lora_rank as configurable parameter to Python SDK