Orchestra Research

This skill helps you accelerate RL-based LLM post-training with slime's Megatron-LM and SGLang for scalable data generation and rollout.

This skill guides researchers through structured ideation frameworks to uncover high-impact research directions, offering actionable prompts and evaluation

This skill helps researchers generate genuinely novel CS and AI ideas by applying cognitive science frameworks like combinatorial creativity and constraint

This skill provides expert guidance for distributed training with DeepSpeed, covering ZeRO, pipeline parallelism, FP16/BF16/FP8, and optimization best

This skill detects prompt injections and jailbreak attempts in LLM apps, ensuring safer interactions and reliable third-party data filtering.

This skill helps you integrate PyTorch FSDP2 into training scripts with correct initialization, sharding, mixed precision, and DTensor-based checkpointing.

This skill speeds up structured generation and agent workflows with prefix caching for JSON outputs and tool calls, boosting latency and throughput.

This skill helps you accelerate RL experimentation with PyTorch-native abstractions, enabling clean algorithm isolation and scalable training using torchforge.

This skill provides expert guidance for fast fine-tuning with Unsloth, enabling 2-5x training speed and reduced memory usage.

This skill guides reinforcement learning based training of large language models using verl across PPO, GRPO, and other RL algorithms.