firecrawl/ai-research-skills
83 skills1k installs1.4k starsGitHub
Install
npx skills add https://github.com/firecrawl/ai-research-skillsSkills in this repo
1Llama CppThe llama-cpp skill documents llama.cpp pure C++ inference with brew or make builds, Metal CUDA and ROCm flags, GGUF downloads from HuggingFace, llama-cli interactive chat, and llama-server OpenAI-compatible HTTP on port 8080. Quantization table compares Q4_K_M recommended default, Q5_K_M quality, and size tradeoffs for 7B models. Use on CPU-only, M1-M4 Macs, AMD Intel GPUs, Raspberry Pi edge, or when avoiding NVIDIA CUDA stacks. Alternatives note TensorRT-LLM for datacenter NVIDIA throughput and vLLM for Python NVIDIA serving with PagedAttention.16installs2Audiocraft Audio Generationaudiocraft-audio-generation covers Meta AudioCraft for text-to-music with MusicGen and text-to-sound with AudioGen including melody-conditioned generation workflows on PyTorch.14installs3Ml Paper WritingThe ml-paper-writing skill guides publication-ready drafting for NeurIPS, ICML, ICLR, ACL, AAAI, and COLM combining researcher writing philosophy with LaTeX templates, reviewer guidelines, and citation verification via Semantic Scholar and arXiv APIs. Core workflow explores research repos, delivers proactive full first drafts when contribution is clear, searches literature programmatically, and iterates from scientist feedback. Critical rule forbids hallucinated citations: never write BibTeX from memory, always fetch programmatically, mark [CITATION NEEDED] when unverified, noting ~40% AI citation error rates. Includes conference checklists and camera-ready formatting guidance.14installs4SglangThe sglang skill covers SGLang inference serving with RadixAttention prefix caching for faster agentic and few-shot workloads, structured JSON and regex outputs, and multi-GPU tensor parallel launch_server commands. Installation uses pip install sglang[all] with optional FlashInfer wheels. Quick start includes launch_server for Llama models, sgl.function decorators, sgl.gen with regex constraints for JSON extraction, and comparisons noting vLLM for plain generation and TensorRT-LLM for NVIDIA-only max latency cases. Documents 5x faster agentic prefix reuse and 3x faster JSON decoding claims from upstream docs. Use for production-scale structured generation and tool-calling agent backends.14installs5Stable Diffusion Image Generationstable-diffusion-image-generation is a Claude Code skill in the AI & Agent Building category.14installs6UnslothThe unsloth skill provides expert guidance for fast fine-tuning using Unsloth with 2-5x faster training and 50-80% less memory via LoRA and QLoRA on Llama, Mistral, Gemma, and Qwen families. It triggers when working with Unsloth features, APIs, debugging, or best practices and references bundled references/llms-txt.md for deeper documentation. Workflow covers beginners starting with getting_started references, feature-specific api and guides files, and scripts or assets folders for automation templates. Notes indicate the skill was auto-generated from official docs with refresh via re-scraper configuration.14installs7ChromaThe chroma skill documents Chroma vector storage for embeddings with metadata, vector search, full-text search, and simple Python or JS APIs scaling from notebooks to clusters. Use when building semantic search, RAG, or document retrieval with self-hosted Apache 2.0 stack. Quick start covers pip install chromadb or npm chromadb plus default embed package, collection create, add documents, and query with where metadata filters. Alternatives table positions Pinecone for managed cloud, FAISS for pure similarity, Weaviate and Qdrant for other production profiles. Metrics cite GitHub stars and weekly release cadence from upstream docs.13installs8Crewai Multi AgentThe crewai-multi-agent skill covers CrewAI teams of autonomous agents with roles, goals, backstories, sequential or hierarchical processes, short and long-term memory, and 50+ tools without LangChain dependencies. Quick start installs crewai and crewai[tools], defines Agent and Task objects, assembles Crew with process mode, and kicks off workflows. Dual paradigm includes Crews for autonomous collaboration and Flows for event-driven control. Use when building researcher-writer-analyst pipelines, production workflows needing tracing, or simpler multi-agent setup than LangGraph. Alternatives note LangChain for general LLM apps, LangGraph for cyclic state, AutoGen for Microsoft chat patterns.13installs9DspyThe dspy skill teaches Stanford DSPy framework for programming rather than prompting LLM systems using Signatures for typed inputs outputs, Modules like Predict and ChainOfThought, and optimizers teleprompters that tune prompts and weights from data. Quick start configures dspy.Claude or other LMs, defines QA Signature classes with InputField and OutputField, and runs Predict or ChainOfThought modules. Core concepts cover inline versus class signatures, composable modules, and metric-driven optimization loops. Use when building reliable NLP pipelines, optimizing prompts from examples, or replacing brittle prompt strings with structured programs.13installs10Evaluating Code ModelsThe evaluating-code-models skill runs bigcode-evaluation-harness across HumanEval, HumanEval+, MBPP, MBPP+, MultiPL-E eighteen languages, APPS, and DS-1000 benchmarks with pass@k metrics. Installation clones bigcode-project repo, pip install -e ., and accelerate config. Evaluation uses accelerate launch main.py with model id, tasks flag, max_length_generation, temperature, n_samples, batch_size, and allow_code_execution for sandboxed scoring. Workflow checklist covers benchmark suite selection, generation config, execution, and pass@k analysis. Industry standard from BigCode used by HuggingFace leaderboards for StarCoder and CodeLlama comparisons.13installs11Evaluating Llms HarnessThe evaluating-llms-harness skill uses EleutherAI lm_eval CLI to score models on MMLU, GSM8K, HellaSwag, TruthfulQA, ARC, HumanEval, and MBPP with standardized few-shot settings. Install via pip install lm-eval. HuggingFace models run with --model hf --model_args pretrained=... --tasks mmlu,gsm8k --device cuda:0 --batch_size auto including 4-bit load_in_4bit options. Standard release suite combines mmlu, gsm8k, hellaswag, truthfulqa, arc_challenge. Supports custom checkpoint paths and API model backends. Industry standard for academic model quality reporting used by labs and HuggingFace.13installs12Huggingface TokenizersThe huggingface-tokenizers skill documents tokenizers library with sub-20 second per GB CPU tokenization, BPE WordPiece and Unigram algorithms, custom vocabulary training, and character alignment maps back to source text. Quick start loads pretrained bert-base-uncased, encodes to tokens and ids, trains custom BPE with BpeTrainer vocab_size and special tokens on corpus files, and saves JSON tokenizer configs. Use for production NLP pipelines, large corpus preprocessing, or custom tokenizer authoring. Alternatives mention SentencePiece for T5, tiktoken for GPT BPE, and AutoTokenizer for pretrained-only loading.13installs13LangchainThe langchain skill covers LangChain framework for agents, chains, tool calling, memory, and vector retrieval across OpenAI, Anthropic, and Google providers with 500+ integrations. Quick start installs langchain plus langchain-openai or langchain-anthropic and langchain-community extras. Basic usage invokes ChatAnthropic models, create_agent builds ReAct agents with tools in under ten lines, and RAG patterns connect vector stores like langchain-chroma. Metrics cite GitHub stars and contributor scale from docs. Alternatives position LlamaIndex for document Q&A, LangGraph for stateful workflows, Haystack for search pipelines.13installs14Langsmith ObservabilityThe langsmith-observability skill configures LangSmith for debugging prompts chains and agents, systematic dataset evaluation, production monitoring, and regression testing. Install langsmith, set LANGSMITH_API_KEY and LANGSMITH_TRACING=true, then use @traceable decorator or wrap_openai for automatic run capture forming hierarchical traces. Core concepts distinguish runs, traces, datasets built from production examples, and custom evaluators. Integrates with OpenAI, Anthropic, LangChain, and LlamaIndex. Alternatives compare Weights and Biases for training tracking and MLflow for model registry focus.13installs15Long ContextThe long-context skill covers techniques to process 32k through 128k+ token documents by extending pretrained LLaMA and Mistral models using Rotary Position Embeddings, YaRN scaling, ALiBi attention biases, and position interpolation fine-tuning. Use when deploying long document QA, extending context beyond original training length, or implementing efficient positional encodings with flash-attn dependencies. Sections explain when to apply each method, training versus inference extrapolation strategies, and fine-tuning existing checkpoints with minimal compute. Pairs with transformers and torch ecosystem tooling.13installs16Modal Serverless GpuThe modal-serverless-gpu skill documents Modal serverless GPU platform for launching Python functions with @app.function decorators, GPU types T4 A10 A100, image builds, secrets, and scheduled jobs without managing clusters. Use when needing elastic GPU bursts, fine-tuning jobs, batch inference, or sandboxes replacing manual CUDA server provisioning. Quick start covers modal pip install, modal token set, stub definition, modal run CLI, and deployment to persistent apps. Patterns include parallel map over datasets, volume mounts for model weights, and web endpoint exposure.13installs17Optimizing Attention FlashThe optimizing-attention-flash skill explains Flash Attention IO-aware exact attention reducing memory from quadratic to linear in sequence length, enabling longer contexts and faster training. Workflow checks PyTorch 2.2+, replaces manual softmax attention with torch.nn.functional.scaled_dot_product_attention, and forces flash backend via torch.backends.cuda.sdp_kernel enable_flash True. Optional flash-attn package provides flash_attn_func for custom models. Includes profiling with torch.utils.benchmark to verify speedup and accuracy parity against baseline attention. Use when training or serving transformers with long sequences on NVIDIA GPUs.13installs18PineconeThe pinecone skill documents Pinecone serverless vector indexes with upsert, query, metadata filtering, namespaces, and hybrid dense plus sparse search targeting sub-100ms p95 latency and auto-scaling to billions of vectors. Python client create_index uses ServerlessSpec on aws gcp or azure regions with cosine euclidean or dotproduct metrics. Core operations cover upsert with metadata, top_k query with include_metadata, and PodSpec alternative for consistent performance tiers. Use for managed production RAG without self-hosting infrastructure. Alternatives cite Chroma self-hosted, FAISS offline, and Weaviate feature-rich self-host.13installs19Prompt GuardThe prompt-guard skill runs meta-llama/Prompt-Guard-86M transformers classifier labeling BENIGN, INJECTION embedded instructions in data, and JAILBREAK direct override attempts with 99%+ true positive and under 1% false positive claims and sub-2ms GPU latency. get_jailbreak_score softmaxes label 2 probability for user input filtering before LLM calls. Workflows cover user prompt filtering with thresholds, third-party RAG data sanitization, and batch processing for document ingestion security. Multilingual support across eight languages. Deploy via HuggingFace pipelines for LLM app input and retrieval guardrails.13installs20Ray TrainThe ray-train skill uses ray[train] TorchTrainer wrapping train_func with ScalingConfig num_workers and use_gpu for distributed coordination, fault tolerance, checkpointing, and metric aggregation. Minimal changes add train.torch.prepare_model and prepare_data_loader inside training loop and train.report for metrics. Scales from laptop multi-GPU to multi-node clusters and integrates Ray Tune hyperparameter sweeps. Handles GPU allocation, worker restarts, and elastic scaling without manual process group setup. Supports HuggingFace, PyTorch, and TensorFlow training functions.13installs21Serving Llms Vllmserving-llms-vllm guides vLLM for high-throughput LLM inference with PagedAttention block KV cache and continuous batching delivering up to 24x throughput over standard transformers. Covers pip install, offline LLM.generate, vllm serve OpenAI-compatible endpoints, GPTQ/AWQ/FP8 quantization, and tensor parallelism for multi-GPU NVIDIA deployment.13installs22Tensorrt Llmtensorrt-llm documents NVIDIA TensorRT-LLM for maximum LLM inference on A100, H100, and GB200 GPUs. Docker-first setup, engine compilation, FP8/INT4/FP4 quantization, in-flight batching, and multi-node scaling target 24,000+ tokens per second. Includes comparison guidance versus vLLM and llama.cpp for hardware-specific choices.13installs23Torchforge Rl Trainingtorchforge-rl-training covers Meta torchforge for PyTorch-native agentic RL separating algorithms from Monarch infrastructure. Integrates TorchTitan training, vLLM inference, and TorchStore sync with GRPO, DAPO, CISPO, GSPO, and SAPO losses in minimal code. Scales single GPU to large clusters without Ray dependency.13installs24Autogpt Agentsautogpt-agents guides the Autogpt platform for building visual workflow agents and persistent autonomous automation. Covers multi-step AI automation design, deployment of continuous agents, and complex orchestration beyond single-shot prompts.12installs25Awq Quantizationawq-quantization documents Activation-aware Weight Quantization for 4-bit LLM compression with roughly 3x speedup and minimal accuracy loss. Targets 7B to 70B deploys on limited GPU memory with better accuracy preservation than GPTQ for instruction-tuned and multimodal models.12installs26Axolotlaxolotl provides YAML-driven LLM fine-tuning for 100+ models with LoRA, QLoRA, DPO, KTO, ORPO, and GRPO plus multimodal support. Covers config templates, training launch, and evaluation for efficient fine-tuning workflows.12installs27Blip 2 Vision Languageblip-2-vision-language explains BLIP-2 bridging frozen image encoders and LLMs for captioning, VQA, image-text retrieval, and multimodal chat with strong zero-shot performance.12installs28Clipclip covers OpenAI CLIP for zero-shot image classification, image-text matching, and cross-modal retrieval trained on 400M image-text pairs. Use for image search, moderation, and vision-language tasks without task-specific fine-tuning.12installs29Constitutional Aiconstitutional-ai documents Anthropic Constitutional AI with self-critique supervised phases and RLAIF alignment. Reduces harmful outputs without human preference labels and underpins Claude safety approaches.12installs30Deepspeeddeepspeed guides Microsoft DeepSpeed distributed training with ZeRO stages, pipeline parallelism, FP16/BF16/FP8 mixed precision, MoE training, and DeepNVMe I/O optimizations for large-scale model training.12installs31Distributed Llm Pretraining Torchtitandistributed-llm-pretraining-torchtitan covers PyTorch torchtitan pretraining with FSDP2, tensor, pipeline, and context parallelism from 8 to 512+ GPUs. Supports Llama 3.1, DeepSeek V3, Float8, torch.compile, and distributed checkpointing.12installs32Faissfaiss documents Facebook AI Similarity Search for billion-scale dense vector k-NN with GPU acceleration. Covers Flat, IVF, and HNSW index types for fast similarity search and clustering without metadata filtering.12installs33Fine Tuning With Trlfine-tuning-with-trl guides HuggingFace TRL for SFT, DPO, PPO, GRPO, and reward model training. Aligns models with human or AI preferences using transformers-native RLHF workflows.12installs34Gguf Quantizationgguf-quantization explains GGUF format and llama.cpp quantization for 2 to 8 bit CPU and Apple Silicon inference without requiring datacenter GPUs.12installs35Gptqgptq covers post-training 4-bit GPTQ quantization with about 4x memory reduction and under 2% perplexity degradation for 70B class models on consumer GPUs with transformers and PEFT QLoRA integration.12installs36Grpo Rl Traininggrpo-rl-training provides TRL-based GRPO and RL fine-tuning guidance for reasoning and task-specific model training with reward shaping and policy optimization patterns.12installs37Guidanceguidance documents Microsoft Guidance for constrained LLM generation with regex, grammars, and guaranteed valid JSON, XML, or code outputs in multi-step workflows.12installs38Hqq Quantizationhqq-quantization explains Half-Quadratic Quantization to 4, 3, or 2 bit without calibration datasets for fast quantization compatible with vLLM and HuggingFace Transformers.12installs39Huggingface Acceleratehuggingface-accelerate covers the unified Accelerate API adding distributed DeepSpeed, FSDP, Megatron, or DDP in a few lines with automatic device placement and FP16, BF16, or FP8 mixed precision.12installs40Implementing Llms Litgptimplementing-llms-litgpt guides Lightning LitGPT implementations for 20+ architectures including Llama, Gemma, Phi, Qwen, and Mistral with single-file educational code and LoRA or QLoRA fine-tuning.12installs41Instructorinstructor documents structured LLM outputs with Pydantic validation, automatic retries, type-safe JSON parsing, and streaming partial results for production extraction pipelines.12installs42Knowledge Distillationknowledge-distillation covers teacher-to-student LLM compression with temperature scaling, soft targets, reverse KLD, and logit distillation to retain performance in smaller deployable models.12installs43Lambda Labs Gpu Cloudlambda-labs-gpu-cloud explains Lambda reserved and on-demand GPU instances with SSH access, persistent filesystems, and multi-node clusters for training and inference workloads.12installs44Llama Factoryllama-factory guides LLaMA-Factory WebUI and CLI fine-tuning for 100+ models with 2 through 8 bit QLoRA and multimodal support.12installs45Llamaguardllamaguard documents Meta LlamaGuard 7-8B moderation across six safety categories with 94-95% accuracy for input and output filtering via vLLM, HuggingFace, or SageMaker.12installs46Llamaindexllamaindex covers the data framework for RAG with 300+ connectors, vector indices, query engines, agents, and multimodal ingestion for document Q&A and enterprise knowledge retrieval.12installs47Llavallava explains Large Language and Vision Assistant combining CLIP encoders with Vicuna or LLaMA for multi-turn image chat, VQA, and visual instruction following.12installs48Mamba Architecturemamba-architecture documents selective state-space Mamba models with O(n) sequence complexity, million-token context, and 5x faster inference than quadratic attention without KV cache.12installs49Miles Rl Trainingmiles-rl-training guides enterprise RL training with miles, a production slime fork for large MoE FP8 or INT4 training with train-inference alignment and speculative RL throughput.12installs50Mlflowmlflow covers experiment tracking, model registry versioning, reproducible runs, and production deployment across frameworks with MLflow's lifecycle APIs.12installs51Model Mergingmodel-merging explains mergekit blending fine-tuned models via SLERP, TIES, and related methods to combine domain skills without full retraining.12installs52Model Pruningmodel-pruning documents Wanda, SparseGPT, and structured pruning to reach about 50% sparsity with minimal accuracy loss for faster inference.12installs53Moe Trainingmoe-training guides Mixture of Experts training with DeepSpeed or HuggingFace for sparse models like Mixtral 8x7B with roughly 5x cost reduction versus dense equivalents.12installs54Nanogptnanogpt presents Andrej Karpathy's educational GPT in about 300 lines reproducing GPT-2 124M on OpenWebText for learning transformer internals from scratch.12installs55Nemo Curatornemo-curator covers NVIDIA GPU-accelerated dataset curation with fuzzy dedup, 30+ quality heuristics, semantic dedup, PII redaction, and NSFW detection at scale.12installs56Nemo Evaluator Sdknemo-evaluator-sdk documents NVIDIA evaluation across 100+ benchmarks and 18+ harnesses with Docker, Slurm, or cloud backends for reproducible LLM and VLM evals.12installs57Nemo Guardrailsnemo-guardrails explains NVIDIA runtime guardrails with Colang 2.0 for jailbreak detection, hallucination checks, PII filtering, and toxicity controls on T4-class GPUs.12installs58Nnsight Remote Interpretabilitynnsight-remote-interpretability guides nnsight activation tracing and interventions with optional NDIF remote execution for 70B+ models without local GPU memory.12installs59Openrlhf Trainingopenrlhf-training covers OpenRLHF with Ray and vLLM for PPO, GRPO, RLOO, and DPO on 7B to 70B models using ZeRO-3 and shared GPU resources.12installs60Outlinesoutlines documents dottxt Outlines for grammar-constrained generation with Pydantic models on local Transformers or vLLM backends.12installs61Peft Fine Tuningpeft-fine-tuning guides HuggingFace PEFT with LoRA, QLoRA, and 25+ parameter-efficient methods training under 1% of weights for 7B to 70B models.12installs62Phoenix Observabilityphoenix-observability covers Arize Phoenix open-source tracing, eval datasets, and production monitoring for debugging LLM application failures.12installs63Pytorch Fsdp2pytorch-fsdp2 adds fully_shard FSDP2 with DeviceMesh, mixed precision, offload, and distributed checkpointing when models exceed single-GPU memory.12installs64Pytorch Lightningpytorch-lightning documents the Trainer abstraction with DDP, FSDP, DeepSpeed, callbacks, and minimal boilerplate scaling laptop to cluster.12installs65Pyvene Interventionspyvene-interventions guides declarative causal interventions, activation patching, and interchange training for testing hypotheses about model internals.12installs66Qdrant Vector Searchqdrant-vector-search explains Qdrant Rust-powered vector search with hybrid filtering for production RAG and semantic retrieval at scale.12installs67Quantizing Models Bitsandbytesquantizing-models-bitsandbytes covers 8-bit and 4-bit NF4 or FP4 quantization with QLoRA training and 8-bit optimizers for 50-75% VRAM savings in transformers.12installs68Ray Dataray-data documents Ray Data streaming ETL across CPU and GPU for Parquet, CSV, JSON, and image batches integrated with Ray Train and PyTorch.12installs69Rwkv Architecturerwkv-architecture explains RWKV linear-time RNN-transformer hybrids with infinite context inference and GPT-parallel training without KV cache.12installs70Segment Anything Modelsegment-anything-model covers Meta SAM foundation segmentation with point, box, or mask prompts and automatic mask generation for any object.12installs71Sentencepiecesentencepiece documents language-independent BPE and Unigram tokenization used by T5, ALBERT, and mBART for multilingual and CJK text.12installs72Sentence Transformerssentence-transformers guides 5000+ embedding models for semantic similarity, clustering, retrieval, and multimodal RAG vectorization.12installs73Simpo Trainingsimpo-training explains Simple Preference Optimization as a reference-free DPO alternative with reported AlpacaEval 2.0 gains and simpler training.12installs74Skypilot Multi Cloud Orchestrationskypilot-multi-cloud-orchestration covers SkyPilot job orchestration across clouds with spot recovery and cost-optimized GPU selection.12installs75Slime Rl Trainingslime-rl-training guides Megatron plus SGLang RL post-training with slime for GLM models and custom data generation at RL scale.12installs76Sparse Autoencoder Trainingsparse-autoencoder-training documents SAELens training and analysis to decompose activations into interpretable monosemantic features.12installs77Speculative Decodingspeculative-decoding covers draft-model, Medusa, and lookahead decoding for 1.5 to 3.6x inference speedups with latency reduction techniques.12installs78Tensorboardtensorboard guides TensorFlow and PyTorch experiment visualization with scalar metrics, histograms, embeddings, and hyperparameter comparison. Covers logdir setup, tensorboard launch, custom scalar and image logging, and integration with training loops for real-time loss and accuracy monitoring.12installs79Training Llms Megatrontraining-llms-megatron documents NVIDIA Megatron-LM for large-scale transformer pretraining with tensor, pipeline, and sequence parallelism across hundreds of GPUs. Covers distributed optimizer states, mixed precision, checkpointing, and scaling laws for billion-parameter models.12installs80Transformer Lens Interpretabilitytransformer-lens-interpretability explains Neel Nanda TransformerLens for mechanistic interpretability with activation caching, hook points, and attribution patching on GPT-2 class models without custom forward passes.12installs81Verl Rl Trainingverl-rl-training covers ByteDance verl for scalable RLHF and GRPO training with Ray, vLLM rollouts, and hybrid engine placement for 7B to 70B policy models.12installs82Weights And Biasesweights-and-biases documents W&B experiment tracking with run logging, hyperparameter sweeps, artifact versioning, and team dashboards for ML training reproducibility.12installs83Whisperwhisper is a Claude Code skill for ai & agent building. It helps solo builders move faster with AI-assisted development.12installs