AI Research Skills Library

The AI Research Skills Library is an open-source collection of 98 specialized skills designed to enable AI agents to autonomously conduct the full AI research lifecycle. This includes everything from initial idea generation and literature surveys to experiment execution (training, evaluation, deployment) and final paper writing.

The library provides two main layers:

1. Research Orchestration Layer: Skills like autoresearch, ideation, and paper writing that manage the high-level research workflow.
2. Engineering Skills: Deep, production-ready domain expertise for specific frameworks (e.g., Megatron-LM, vLLM, TRL) covering areas like distributed training, inference, and fine-tuning.

Unsloth is an open-source framework for LLM fine-tuning and reinforcement learning (RL). It is designed to train, run, evaluate, and save models like gpt-oss, Llama, DeepSeek, TTS, Qwen, Mistral, and Gemma up to 2x faster with 70% less VRAM.

Key Capabilities:

• Training Modes: Supports full-finetuning, pretraining, 4-bit, 16-bit, and 8-bit training.
• Model Support: Supports vision, text-to-speech (TTS), BERT, and multimodal models. Any model compatible with transformers works in Unsloth.
• Reinforcement Learning: Highly efficient RL library using 80% less VRAM, supporting GRPO, GSPO, etc.
• Hardware Support: Linux, Windows, Colab, Kaggle, NVIDIA, AMD, and Intel.

Unsloth Dynamic 2.0 is an advanced quantization method for GGUFs that outperforms standard methods in 5-shot MMLU and KL Divergence.

Key improvements in v2.0 include:

• Intelligent Layer Selection: Dynamically adjusts quantization types for every layer individually.
• Expanded Model Support: Works on all architectures, including both MoE (Mixture of Experts) and non-MoE models.
• Custom-Tailored Schemes: Each model (e.g., Gemma 3 vs Llama 4) uses a unique quantization scheme optimized for its specific architecture.
• New Formats: Added support for Q4_NL, Q5.1, Q5.0, Q4.1, and Q4.0 to maximize efficiency on Apple Silicon and ARM devices.
• Improved Calibration: Uses a high-quality dataset of >1.5M tokens to enhance conversational chat performance.

DeepSpeed is a library designed for extreme speed and scale in Deep Learning (DL) training. It provides system innovations to make large-scale training effective, efficient, and easier to use. Key technologies include:

• ZeRO (Zero Redundancy Optimizer): Optimizes memory usage.
• 3D-Parallelism: Combines data, pipeline, and tensor parallelism.
• DeepSpeed-MoE: Support for Mixture-of-Experts models.
• ZeRO-Infinity: Enables training of massive models by offloading to CPU/NVMe.
• SuperOffload: Large-scale LLM training on superchips.
• ZenFlow: A stall-free offloading engine for LLM training.
• Arctic Long Sequence Training (ALST): Scalable training for multi-million token sequences.
• DeepNVMe: Affordable I/O scaling for DL applications.

Qwen3-VL is a vision model series from Qwen featuring instruct and thinking versions.

• Model Sizes: 2B, 4B, 8B, and 32B (dense); 30B and 235B (MoE).
• Capabilities: Vision, video, and OCR.
• Context Window: 256K (extendable to 1M).
• Local Execution: Supported via GGUF in llama.cpp (as of Oct 30, 2025).
• Unsloth Support: Supports fine-tuning and Reinforcement Learning (RL) for Qwen3-VL.

Unsloth provides tools to transform standard models into reasoning models using Reinforcement Learning (RL).

Key Capabilities:

• Model Scale: With 15GB VRAM, you can transform models up to 17B parameters (e.g., Llama 3.1 8B, Phi-4 14B, Mistral 7B, or Qwen2.5 7B).
• Vision/Multimodal Support: Unsloth supports RL for Vision/multimodal (VLM) models.
• Low VRAM Training: A minimum of 5GB VRAM is sufficient to train reasoning models for any model with 1.5B parameters or less.

Important Note: RL is most effective when starting from an already instruction-finetuned model. If the probability of a correct response is always 0, RL will not work; instruction tuning helps boost that probability above 0.

OpenVLA-OFT replaces standard discrete action tokenization with continuous action heads (L1 regression or diffusion) while keeping the VLA backbone frozen and using LoRA for adaptation.

Variants:

• OFT: Uses 2 images (front + wrist) and no FiLM conditioning. Typically used for LIBERO simulation.
• OFT+: Uses 3 images (high + left + right wrist) and FiLM (Feature-wise Linear Modulation) conditioning. Typically used for ALOHA real-world setups.

Key Components:

• LoRA adaptation: Rank-32 LoRA on the VLA backbone.
• Continuous actions: Supports L1 regression (default) or diffusion heads.
• Multi-image input: Configurable via num_images_in_input.

To avoid duplication, the CLI uses a Canonical Storage strategy.

1. Skills are downloaded to a single source of truth: ~/.orchestra-skills/.
2. A .lock.json file in the canonical directory manages versioning and agent mapping.
3. The CLI creates symlinks in each agent's specific skills directory (e.g., ~/.claude/skills/) that point back to the files in ~/.orchestra-skills/.

This ensures that updating a skill once in the canonical storage updates it for all agents using that skill.

Agents in AutoGPT are represented as graphs consisting of nodes connected by links. Each node contains a Block, which is a reusable functional component.

Block Types:

• INPUT: Agent entry points
• OUTPUT: Agent outputs
• AI: LLM calls, text generation
• WEBHOOK: External triggers
• STANDARD: General operations
• AGENT: Nested agent execution

The llama-factory skill provides expert guidance for fine-tuning Large Language Models (LLMs) using the LLaMA-Factory framework. It supports over 100 models, no-code WebUI interfaces, various quantization levels (2/3/4/5/6/8-bit QLoRA), and multimodal support.

Trigger this skill when:

• Working with llama-factory development.
• Inquiring about llama-factory features or APIs.
• Implementing llama-factory solutions.
• Debugging llama-factory code.
• Learning llama-factory best practices.