Gradient

XGBoost is a scalable gradient-boosting software library for regression, classification, ranking, and user-defined objectives. It supports multiple languages, including C++, Python, R, Java, Scala, and Julia, and runs on Windows, Linux, macOS, and major cloud platforms. Its optimized backend supports distributed training across multiple machines and integrations with systems such as Spark and Dask, with the project describing use at beyond-billion-example scale. In the cited fraud-detection system, XGBoost serves as the tabular classifier and receives GraphSAGE-generated graph embeddings as additional features.

rLLM is an open-source framework from UC Berkeley's Sky Computing Lab for training language agents with reinforcement learning. It connects agent code to evaluation and training through a decorator-based Python API or CLI workflow: agents run in sandboxes, the model gateway captures token IDs and log probabilities, traces are organized into episodes and trajectories, evaluators compute rewards, and a training backend updates the model. It supports existing or custom agent harnesses, Docker, Daytona, Modal, and local sandboxes, with training backends including verl, tinker, and Fireworks, and methods including GRPO, REINFORCE, RLOO, supervised fine-tuning, and on-policy distillation. The project also provides integrated benchmarks and requires Python 3.11 or later.

Miles is an enterprise-facing reinforcement learning framework for large-scale post-training of large language and vision-language models. It pairs SGLang for high-throughput, agentic rollouts with Megatron-LM for scalable training, and also provides a PyTorch FSDP2 backend for Hugging Face implementations. Its asynchronous architecture decouples rollout and training workers, supports configurable on- and off-policy schedules, and updates rollout engines in-loop through peer-to-peer RDMA weight transfer. The framework includes token-in-token-out data flow, Rollout Routing Replay for replaying mixture-of-experts routing decisions during training, fault-tolerant recovery of failed SGLang engines, low-precision training with MXFP8 and NVFP4 alongside FP8, INT4 QAT, BF16, and FP16, and LoRA or multi-LoRA training. It supports reinforcement-learning recipes including GRPO, GSPO, PPO, and REINFORCE++, as well as supervised fine-tuning, on-policy distillation, agentic environments, and diffusion-model training. Miles was forked from slime and integrates SGLang, Megatron-LM, and torch_memory_saver; the repository is released as an open-source project, though the provided page text does not state its license.

AutoBrowse is a browser-agent learning loop that improves task-specific behavior without changing model weights. It uses two Claude Agent SDK loops: a doer agent operates the browser, while a teacher agent reviews screenshots, traces, and task results after each run and writes lessons to a strategy.md file for the next attempt. When the strategy converges, the loop produces a reusable skill.md file containing site-specific navigation choices, waiting behavior, and other learned procedures that future agents can apply. The approach depends on replayable environments and verifiable outcomes, and the videos describe monitoring and failure-triggered repair loops for changing websites, with generated skills requiring review because of prompt-injection risks.

Megatron is a training framework that handles data-parallel ranks and model-weight optimization. In the described multimodal training pipeline, it acts as a downstream consumer of the data pipeline rather than managing the data-loading bottlenecks.