Literature, experiments, and deep research workflows.
Deep research workflows that go beyond one search query: literature review and citation handling, experiment tracking, scientific computing, and multi-source synthesis with the sourcing kept intact. These artifacts are built for questions where the answer needs evidence attached, and where reading twenty sources properly beats skimming three.
Ranked by GitHub stars. Search to find fast, or page through the full list.
Expert guidance for fine-tuning LLMs with Axolotl - YAML configs, 100+ models, LoRA/QLoRA, DPO/KTO/ORPO/GRPO, multimodal support
Language-independent tokenizer treating text as raw Unicode. Supports BPE and Unigram algorithms. Fast (50k sentences/sec), lightweight (6MB memory), deterministic vocabulary. Used by T5, ALBERT, XLNet, mBART. Train on raw text without pre-tokenization. Use when you need multilingual support, CJK languages, or reproducible tokenization.
CoCounsel Legal delivers comprehensive Westlaw Deep Research reports with inline, linked citations to Westlaw and Practical Law sources.
State-space model with O(n) complexity vs Transformers' O(n²). 5× faster inference, million-token sequences, no KV cache. Selective SSM with hardware-aware design. Mamba-1 (dstate=16) and Mamba-2 (dstate=128, multi-head). Models 130M-2.8B on HuggingFace.
High-performance RLHF framework with Ray+vLLM acceleration. Use for PPO, GRPO, RLOO, DPO training of large models (7B-70B+). Built on Ray, vLLM, ZeRO-3. 2× faster than DeepSpeedChat with distributed architecture and GPU resource sharing.
Triages proposed AI use cases against your registry, runs impact assessments across the regimes in scope, reviews vendor AI terms for training-on-data and liability gaps, and keeps your AI policy current with practice.
Provides PyTorch-native distributed LLM pretraining using torchtitan with 4D parallelism (FSDP2, TP, PP, CP). Use when pretraining Llama 3.1, DeepSeek V3, or custom models at scale from 8 to 512+ GPUs with Float8, torch.compile, and distributed checkpointing.
Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO for preference alignment, PPO/GRPO for reward optimization, and reward model training. Use when need RLHF, align model with preferences, or train from human feedback. Works with HuggingFace Transformers.
Deep research powered by Exa. Use for lead generation, literature reviews, deep dives, competitive analysis, or any query where one search falls short, including phrases like 'research this', 'find everything about', 'find me all', or 'deep dive on'.
Implements and trains LLMs using Lightning AI's LitGPT with 20+ pretrained architectures (Llama, Gemma, Phi, Qwen, Mistral). Use when need clean model implementations, educational understanding of architectures, or production fine-tuning with LoRA/QLoRA. Single-file implementations, no abstraction layers.
Provides guidance for training and analyzing Sparse Autoencoders (SAEs) using SAELens to decompose neural network activations into interpretable features. Use when discovering interpretable features, analyzing superposition, or studying monosemantic representations in language models.
Simulates academic peer review, evaluating papers across Originality, Methodology, Results, and Writing to provide Major/Minor Revision recommendations with actionable feedback. Triggers when a user asks to "review my paper," "simulate peer review," or "give my paper a peer review.
Provides guidance for LLM post-training with RL using slime, a Megatron+SGLang framework. Use when training GLM models, implementing custom data generation workflows, or needing tight Megatron-LM integration for RL scaling.
Provides guidance for PyTorch-native agentic RL using torchforge, Meta's library separating infra from algorithms. Use when you want clean RL abstractions, easy algorithm experimentation, or scalable training with Monarch and TorchTitan.
长篇网文扫榜。分析起点、番茄、晋江等平台排行榜数据，提炼市场趋势与热门题材。触发方式：/story-long-scan、/长篇扫榜、「长篇什么火」「起点排行」。
Expert guidance for GRPO/RL fine-tuning with TRL for reasoning and task-specific model training
Provides guidance for enterprise-grade RL training using miles, a production-ready fork of slime. Use when training large MoE models with FP8/INT4, needing train-inference alignment, or requiring speculative RL for maximum throughput.
短篇网文扫榜。分析知乎盐言、七猫、黑岩、点众等平台热门短篇数据，捕捉风口题材。触发方式：/story-short-scan、/短篇扫榜、「短篇什么火」「知乎故事排行」。
Provides guidance for performing causal interventions on PyTorch models using pyvene's declarative intervention framework. Use when conducting causal tracing, activation patching, interchange intervention training, or testing causal hypotheses about model behavior.
GPU-accelerated data curation for LLM training. Supports text/image/video/audio. Features fuzzy deduplication (16× faster), quality filtering (30+ heuristics), semantic deduplication, PII redaction, NSFW detection. Scales across GPUs with RAPIDS. Use for preparing high-quality training datasets, cleaning web data, or deduplicating large corpora.
External research workflow for docs, web, APIs - NOT codebase exploration
Analyze repository structure, patterns, conventions, and documentation for understanding a new codebase
Cross-CLI skill for Obsidian.
Formal theorem proving with research, testing, and verification phases
Search Mathlib for lemmas by type signature pattern
Convert Kanchi-style dividend investing into a repeatable US-stock operating procedure. Use when users ask for かんち式配当投資, dividend screening, dividend growth quality checks, PERxPBR adaptation for US sectors, pullback limit-order planning, or one-page stock memo creation. Covers screening, deep dive, entry planning, and post-purchase monitoring cadence.
Detect IBD-style Distribution Days for QQQ/SPY (close down at least 0.2% on higher volume), track 25-session expiration and 5% invalidation, count d5/d15/d25 clusters, classify market risk (NORMAL/CAUTION/HIGH/SEVERE), and emit TQQQ/QQQ exposure recommendations. Use after market close, before TQQQ exposure changes, or as input to FTD/market-state frameworks. Does not execute trades.
Generate academic diagrams and statistical plots from text using multi-agent AI.
Abstract detector tickets and hints into reusable edge concepts with thesis, invalidation signals, and strategy playbooks before strategy design/export.
Detects Follow-Through Day (FTD) signals for market bottom confirmation using William O'Neil's methodology. Dual-index tracking (S&P 500 + NASDAQ) with state machine for rally attempt, FTD qualification, and post-FTD health monitoring. Use when user asks about market bottom signals, follow-through days, rally attempts, re-entry timing after corrections, or whether it's safe to increase equity exposure. Complementary to market-top-detector (defensive) - this skill is offensive (bottom confirmation).
Search, download, and read academic papers from 20+ sources (arXiv, PubMed, Semantic Scholar, CrossRef, etc). Use when the user asks to find papers, search for research, look up academic literature, download a paper PDF, or extract text from a paper.
Add items (research objects) to existing research outline.
Summarize deep research results into markdown report, cover all fields, skip uncertain values.
Karpathy: An agentic Machine Learning Engineer
Read research outline, launch independent agent for each item for deep research. Disable task output.
Conduct preliminary research on a topic and generate research outline. For academic research, benchmark research, technology selection, etc.
Simplest distributed training API. 4 lines to add distributed support to any PyTorch script. Unified API for DeepSpeed/FSDP/Megatron/DDP. Automatic device placement, mixed precision (FP16/BF16/FP8). Interactive config, single launch command. HuggingFace ecosystem standard.
Search academic literature and generate research hypotheses
Transcribe audio/video to speaker-labeled text — who-said-what by default, plain-text opt-out; MLX-local on Apple Silicon or remote; local files, media URLs. Use for transcribing recordings/podcasts/lectures/meetings, ASR, speech-to-text, 转录, 语音转文字, 录音转文字, speaker diarization/说话人分离/识别/谁在说话, timestamps 字幕/时间戳/音画对齐, CAM++ voiceprint ID. This skill ALSO owns audio PREPROCESSING for ASR as a first-class trigger, even without transcription: convert any audio/video into an ASR-ready file (转换成适合 ASR 的格式, 转格式, convert/prepare audio for ASR, 音频预处理), downsample to 16kHz mono 16-bit (降采样, 重采样, 单声道, 归一化), merge multi-segment recorder dumps (多段合并/拼接, DJI TX01/TX02), transcode to small M4A + pitch-preserved speedup to cut metered-ASR billed minutes (转 M4A, 压缩上传, 加速, 1.3x, 飞书妙记/Feishu Minutes). Trigger even when it looks like a trivial one-line ffmpeg — the skill owns sample-rate/bit-depth/channel, merge-order, speed-vs-WER, format choices + a blessed prepareasrinput.py.
Minimal AI-driven research engine: idea → literature review → experiment plan → paper draft
Write publication-ready ML/AI/Systems papers for NeurIPS, ICML, ICLR, ACL, AAAI, COLM, OSDI, NSDI, ASPLOS, SOSP. Use when drafting papers from research repos, structuring arguments, verifying citations, or preparing camera-ready submissions. Includes LaTeX templates, reviewer guidelines, and citation verification workflows.
Applies cognitive science frameworks for creative thinking to CS and AI research ideation. Use when seeking genuinely novel research directions by leveraging combinatorial creativity, analogical reasoning, constraint manipulation, and other empirically grounded creative strategies.
Multi-cloud orchestration for ML workloads with automatic cost optimization. Use when you need to run training or batch jobs across multiple clouds, leverage spot instances with auto-recovery, or optimize GPU costs across providers.
Generate a Python code skeleton from an experiment blueprint
Evaluates LLMs across 60+ academic benchmarks (MMLU, HumanEval, GSM8K, TruthfulQA, HellaSwag). Use when benchmarking model quality, comparing models, reporting academic results, or tracking training progress. Industry standard used by EleutherAI, HuggingFace, and major labs. Supports HuggingFace, vLLM, APIs.
Draft a LaTeX research paper from all previous stage outputs
Generates publication-quality figures for ML papers from research context. Given a paper section or description, extracts system components and relationships to generate architecture diagrams via Gemini. Given experiment results or data, auto-selects chart type and generates data-driven figures via matplotlib/seaborn. Use when creating any figure for a conference paper.
Guides researchers through structured ideation frameworks to discover high-impact research directions. Use when exploring new problem spaces, pivoting between projects, or seeking novel angles on existing work.