Customize Beam - Reflection Developer Docs

Adapt Beam to your own data and tasks with parameter-efficient tuning (LoRA and QLoRA), supervised fine-tuning (SFT), or reinforcement learning (RL).

Recipes for each stack are coming soon. They will cover data preparation, training, evaluation against the baseline model, and export.

Integration stacks

Beam is supported in the open-source training stacks below. We’ll add more stacks over time.

Method	Stack	Best for
SFT	NeMo AutoModel	Hugging Face checkpoints with NVIDIA-optimized training, LoRA and full-parameter
SFT	Megatron Bridge	Large-scale full fine-tuning with Megatron-Core parallelism, with conversion to and from Hugging Face format
SFT	Axolotl	Config-driven LoRA, QLoRA, and full fine-tuning with broad dataset and adapter support
RL	Miles	PyTorch-native RL post-training with SGLang rollouts and Megatron-LM or FSDP training, focused on MoE stability and low-precision training
RL	NeMo RL	Scalable RL post-training on NVIDIA hardware, from a single node to large clusters
RL	SkyRL	Modular RL for agentic and multi-turn tasks, with pluggable environments and inference backends

Last modified on October 16, 2026