Skill UI

Pyvene is a declarative framework designed for performing causal interventions on PyTorch neural networks. It allows researchers to conduct advanced experiments such as activation patching, causal tracing (ROME-style), and interchange intervention training (IIT). Use this library when you need to test causal hypotheses about model behavior, deeply interpret model components, or ensure reproducibility in advanced AI research.

torchforge is Meta's advanced PyTorch-native library designed for agentic Reinforcement Learning (RL). It fundamentally separates core RL algorithms from complex distributed infrastructure concerns. This allows researchers to rapidly prototype and experiment with novel RL methods (like GRPO) using clean, PyTorch-native abstractions, while automatically handling large-scale tasks like distributed training across thousands of GPUs using Monarch and TorchTitan.

Get a deep, multi-round, and adversarial critical review of your research work, ideas, or drafts. This skill simulates the perspective of a senior ML reviewer (e.g., NeurIPS/ICML level) using dedicated external backends (Codex or Manual). It is designed to find logical gaps, missing experiments, and narrative weaknesses, providing brutal honesty to elevate your academic contribution.

This tool systematically designs and plans ablation studies crucial for advanced research paper submissions. It simulates the role of a rigorous ML reviewer, guiding the process of isolating the specific contributions of each novel component, testing hyperparameter sensitivities, and addressing potential reviewer questions. The workflow ensures that every experiment has a clear hypothesis, maximizing the scientific impact of the research results.

A persistent, graph-based knowledge base designed to accumulate structured knowledge throughout the entire research lifecycle. It stores and connects diverse entities—including published papers, novel ideas, conducted experiments, and testable claims. Inspired by LLM Wiki patterns, it ensures knowledge compounding, meaning the wiki becomes progressively smarter with every piece of input, rather than requiring re-derivation for every query.

PufferLib is a high-performance, scalable RL framework designed for massive parallel training and vectorized simulation. Featuring PuffeRL (optimized PPO+LSTM) and the PufferEnv API, it supports multi-agent systems, custom environment creation, and achieves industry-leading speeds (up to millions of steps/second). Ideal for advanced research and large-scale AI experimentation.

A comprehensive framework designed to structure and automate PyTorch deep learning workflows. It simplifies boilerplate code by organizing models into LightningModules, automating training loops via the Trainer, and providing robust support for multi-GPU/TPU scaling (DDP, FSDP, DeepSpeed), data handling (DataModules), and best-in-class experiment logging (W&B, MLflow). Ideal for professional, scalable deep learning project development.

Stable Baselines3 is a robust, PyTorch-based library providing reliable implementations of state-of-the-art reinforcement learning algorithms (PPO, SAC, DQN, etc.). It is designed for standard RL experiments, quick prototyping, and developing complex single-agent systems. The library supports training agents, creating custom Gymnasium environments, and running vectorized simulations for accelerated learning.

A robust orchestration skill designed for managing large-scale, complex machine learning experiments on remote GPU clusters via SSH. It handles multi-seed sweeps, sequential wave transitions, and deep learning pipelines requiring state management. Features include OOM-aware automatic retries, cleanup of stale terminal sessions, and deterministic scheduling to prevent race conditions, making it ideal for grid computing and staged training.