What's New
Tinker Integration for Fine-tuning (#710)
- Added support for Tinker SDK integration, enabling LoRA fine-tuning directly after generating data with Curator. When a TINKER_API_KEY is provided, training runs against the Tinker service with real
forward/backward passes. Features include: - Checkpointing support (per-epoch and per-N-steps) with dashboard visibility
- Gradient accumulation, LoRA alpha/dropout/target_modules configuration
- Checkpoint resume support
- Falls back to mock mode when SDK/key is not available
- New examples in examples/tinker/ and examples/poem_finetuning_example.py
Claude 4.x and 3.7 Model Support (#704)
Added multimodal and JSON format support for new Anthropic models:
- Claude 4.5 (Sonnet, Haiku, Opus)
- Claude 4.0/4.1 (Sonnet, Opus)
- Claude 3.7 Sonnet
Bug Fixes
- Fixed LiteLLM multi-hosted_vllm request address mixture (#701) — Resolved an issue where requests could be routed to the wrong vLLM host when using multiple hosted_vllm backends.
- Fixed broken tests (#706)