Changelog

Here are some of the user-facing changes we've made to RunRL recently.

October 11, 2025

LoRA Support: Added comprehensive support for parameter-efficient fine-tuning with LoRA (Low-Rank Adaptation)
- Enables 2-3× faster training with significantly reduced memory usage
- New parameters: use_lora, lora_rank, lora_alpha, lora_dropout
- Automatic MoE detection: Mixture-of-Experts models (e.g., Qwen3-Coder-30B-A3B-Instruct) now automatically use attention-only LoRA for vLLM compatibility
- Python SDK updated to v0.2.2 with full LoRA documentation and examples

Dataset wizard (beta): Easily generate reward functions by uploading a dataset and describing the task you want an LLM to improve at
API parity with the UI: The REST API now covers every user-facing feature—files, runs, deployments, tools, shared configurations, validation, and search.
Docs refresh: Completely overhauled API docs with end-to-end examples and coverage of new endpoints.

SFT warmups: You can now provide a dataset of examples to run SFT on before RL, boosting early performance

Completions display: Improved display for model completions in multi-turn environments