--- title: "Orchestra Research - Air Market" description: "Orchestra Research publishes 68 listings on Air Market." canonical: "https://airmarket.ai/@orchestra-research" --- [Air Market](/) / [Creators](/creators) / Orchestra Research Creator # Orchestra Research **Orchestra Research on Air Market** | | | | --- | --- | | Listings | 68 | | GitHub | [Orchestra-Research](https://github.com/Orchestra-Research) | **Published listings, with what each one does** | Listing | Kind | What it does | | --- | --- | --- | | [verl RL Training](/app/verl-rl-training) | Skill | You have a base model and want to shape its behaviour with reinforcement learning rather than more plain training. This is guidance for ver… | | [torchforge RL Training](/app/torchforge-rl-training) | Skill | You want to experiment with reinforcement learning for agents, but the infrastructure keeps getting tangled up with the algorithm. This is… | | [slime RL Training](/app/slime-rl-training) | Skill | You have a trained model and want to improve it further with reinforcement learning. This guides you through doing that with slime, a frame… | | [SimPO - Simple Preference Optimization](/app/simpo-training) | Skill | Teaching a model which answers people prefer usually means keeping a second reference model alongside it. This method drops that requiremen… | | [RWKV - Receptance Weighted Key Value](/app/rwkv-architecture) | Skill | Transformers get slower and hungrier as the text gets longer, because of the cache they carry. RWKV is a hybrid design that trains in paral… | | [PyTorch FSDP2](/app/pytorch-fsdp2) | Skill | Your model is too big for a single graphics card. This adds sharding to your PyTorch training script, splitting the model across devices, a… | | [OpenRLHF - High-Performance RLHF Training](/app/openrlhf-training) | Skill | Doing RLHF on a large model means coordinating training and generation across a lot of GPUs, which is where most of the difficulty lives. O… | | [Llama-Factory Skill](/app/llama-factory) | Skill | You want to fine-tune a language model but would rather not write the training code. LLaMA-Factory gives you a web interface for it, works… | | [Lambda Labs GPU Cloud](/app/lambda-labs-gpu-cloud) | Skill | You need GPUs for training or inference and do not have them locally. This covers reserved and on-demand GPU cloud instances you connect to… | | [LitGPT - Clean LLM Implementations](/app/implementing-llms-litgpt) | Skill | Sometimes you want to see how a language model is actually put together, without layers of abstraction in the way. This covers implementing… | | [GGUF - Quantization Format for llama.cpp](/app/gguf-quantization) | Skill | You want to run a model but do not have a big GPU, or want it working on your own machine. This covers converting models to the GGUF format… | | [Experiment Tracking With SwanLab](/app/experiment-tracking-swanlab) | Skill | When you train models, it is easy to lose track of which run did what. This helps you record your runs with SwanLab and view them on a dash… | | [Cosmos Policy Evaluation](/app/evaluating-cosmos-policy) | Skill | You have a robot manipulation policy and need to know how it performs before it touches hardware. This sets it up and runs it in simulation… | | [DSPy](/app/dspy) | Skill | You are writing and rewriting prompts by hand and the system is getting hard to keep straight. This lets you describe what each part should… | | [TorchTitan - PyTorch Native Distributed LLM Pretraining](/app/distributed-llm-pretraining-torchtitan) | Skill | You are pretraining a language model and it will not fit on one machine. This uses torchtitan to spread the training across GPUs in several… | | [Deepspeed Skill](/app/deepspeed) | Skill | You are training a model across multiple GPUs or machines and the configuration options are unfamiliar. This gives you guidance on the choi… | | [Constitutional AI - Harmlessness from AI Feedback](/app/constitutional-ai) | Skill | You need a model to stop producing harmful answers, and paying people to label examples is not an option. This is a training approach where… | | [AudioCraft: Audio Generation](/app/audiocraft-audio-generation) | Skill | You need a piece of music or a sound effect and you do not have one. This generates audio from a text description of what you want, either… | | [Weights And Biases](/app/weights-and-biases) | Skill | After enough training runs you cannot remember which settings produced which result. This uses W&B to log experiments automatically, watch… | | [Megatron-Core - Large-Scale LLM Training](/app/training-llms-megatron) | Skill | You are training a model large enough that it will not fit on one GPU, and the question becomes how to spread it out. This covers training… | | [TensorBoard](/app/tensorboard) | Skill | Training a model produces numbers you cannot read as a wall of text. This uses TensorBoard to show them: metric curves, histograms for debu… | | [vLLM - High-Performance LLM Serving](/app/serving-llms-vllm) | Skill | You have a model and now people need to call it, all at once. This serves it with techniques that keep many requests moving through the sam… | | [SentencePiece - Language-Independent Tokenization](/app/sentencepiece) | Skill | Most tokenizers assume text is already split into words, which breaks down for languages that do not put spaces between them. This one read… | | [Sentence Transformers - State-of-the-Art Embeddings](/app/sentence-transformers) | Skill | You want a computer to tell which pieces of text mean roughly the same thing. This turns text - and images - into numbers that can be compa… | | [Ray Train - Distributed Training Orchestration](/app/ray-train) | Skill | Your training job has outgrown one machine. This spreads it across a cluster, from a laptop up to very many nodes, and keeps working with P… | | [bitsandbytes - LLM Quantization](/app/quantizing-models-bitsandbytes) | Skill | The model you want will not fit in your GPU memory. This shrinks it to 8-bit or 4-bit so it does fit, with little cost to accuracy, and wor… | | [Qdrant - Vector Similarity Search Engine](/app/qdrant-vector-search) | Skill | Your app needs to find the items closest in meaning to a query, fast, and over a lot of them. This is a search engine built for that: neare… | | [Phoenix - AI Observability Platform](/app/phoenix-observability) | Skill | Your AI app gave a strange answer and you cannot tell why. This gives you a detailed record of each step it took, so you can look back and… | | [Outlines](/app/outlines) | Skill | Asking a model for JSON and getting something almost-valid back means writing repair code and retries. Outlines constrains generation so th… | | [MLflow](/app/mlflow) | Skill | When you train models repeatedly, it gets hard to remember what settings produced which result, or which version is the live one. MLflow ke… | | [Miles RL Training](/app/miles-rl-training) | Skill | Teaching a very large model through reinforcement learning is fiddly once the model is big and split across experts. This is guidance for d… | | [Long Context](/app/long-context) | Skill | Your documents are longer than the model was built to read, and it stops short. This covers techniques for stretching a transformer's conte… | | [Knowledge Distillation: Compressing LLMs](/app/knowledge-distillation) | Skill | A big model works well but costs too much to run. Knowledge distillation trains a smaller student model to imitate the larger teacher, so y… | | [Instructor](/app/instructor) | Skill | A language model gives you prose when your code needs a typed object, and sometimes the JSON is malformed. This uses Instructor to extract… | | [GPTQ (Generative Pre-trained Transformer Quantization)](/app/gptq) | Skill | You have a model too large for your GPU and no bigger one to hand. This covers four-bit quantization applied after training, which shrinks… | | [TRL - Transformer Reinforcement Learning](/app/fine-tuning-with-trl) | Skill | You have a language model that answers, but not the way you want it to. This covers training it from human feedback using TRL: instruction… | | [OpenPI Fine-Tuning and Serving](/app/fine-tuning-serving-openpi) | Skill | You want to use Physical Intelligence's pi0 robot policy models on your own data or run them for inference. This covers fine-tuning and ser… | | [BigCode Evaluation Harness - Code Model Benchmarking](/app/evaluating-code-models) | Skill | You need to know how good a model actually is at writing code, and opinions are not enough. This runs it against established coding benchma… | | [BLIP-2: Vision-Language Pre-training](/app/blip-2-vision-language) | Skill | You have images and you want a model to describe them, answer questions about them, or match them to text. This connects an existing image… | | [AWQ (Activation-aware Weight Quantization)](/app/awq-quantization) | Skill | You have a large language model and not enough GPU memory to run it comfortably. This compresses the model's weights down to four bits so i… | | [TensorRT-LLM](/app/tensorrt-llm) | Skill | You have a model that works and now it has to serve real traffic on NVIDIA hardware. This covers optimizing inference with TensorRT: quanti… | | [Speculative Decoding](/app/speculative-decoding) | Skill | Your model works but each reply takes too long, and you cannot throw more hardware at it. This covers ways of generating tokens in parallel… | | [Flash Attention - Fast Memory-Efficient Attention](/app/optimizing-attention-flash) | Skill | Transformers get slow and memory-hungry as inputs get longer, and attention is usually the culprit. Flash Attention computes it differently… | | [NeMo Evaluator SDK - Enterprise LLM Benchmarking](/app/nemo-evaluator-sdk) | Skill | Judging a model means running it against a lot of different tests, and stitching those harnesses together yourself is tedious. This runs a… | | [NeMo Curator - GPU-Accelerated Data Curation](/app/nemo-curator) | Skill | Raw scraped data is full of repeats, junk and things that should not be in there. NeMo Curator cleans it up: it finds near-duplicates, filt… | | [MoE Training](/app/moe-training) | Skill | You want a bigger model than your compute allows, because dense models cost more the larger they get. Mixture of Experts activates only par… | | [Model Pruning: Compressing LLMs](/app/model-pruning) | Skill | A model is bigger and slower than you can live with, but retraining it is not on the table. Pruning strips out parts of it in one pass, lea… | | [Model Merging](/app/model-merging) | Skill | You have several fine-tuned models, each good at a different thing, and you want one model that does all of it. Merging combines them direc… | | [Mamba - Selective State Space Models](/app/mamba-architecture) | Skill | Long inputs get expensive fast with the usual transformer setup, and memory grows as you generate. This covers a different kind of sequence… | | [HuggingFace Tokenizers - Fast Tokenization for NLP](/app/huggingface-tokenizers) | Skill | Before a model reads text, the text has to be split into tokens, and on large corpora that step can be slow. This covers fast tokenizers bu… | | [OpenVLA-OFT](/app/fine-tuning-openvla-oft) | Skill | You are working on robot policies with OpenVLA-OFT and need to train, evaluate or deploy one. This walks you through fine-tuning those poli… | | [TransformerLens Interpretability](/app/transformer-lens-interpretability) | Skill | You want to know not just what a model outputs but what it is doing inside to get there. This is guidance for TransformerLens, which lets y… | | [Systems Paper Writing](/app/systems-paper-writing) | Skill | You have systems research to submit and the writing is the hard part. This is a guide to structuring the paper, with blueprints down to the… | | [Stable Diffusion Image Generation](/app/stable-diffusion-image-generation) | Skill | You need pictures and you would rather describe them than draw them. This covers generating images from text with Stable Diffusion, reworki… | | [Sparse Autoencoder Training](/app/sparse-autoencoder-training) | Skill | Inside a language model, many ideas share the same neurons, so nothing is readable on its own. This guides you through training a sparse au… | | [pyvene Interventions](/app/pyvene-interventions) | Skill | You have a hunch about why a model behaves a certain way, and you want to test it rather than argue it. This walks you through intervening… | | [Prompt Guard](/app/prompt-guard) | Skill | People can hide instructions inside text your app feeds to a model, and the model may follow them. This is a detector from Meta that inspec… | | [Conference Talk Slides](/app/presenting-conference-talks) | Skill | Your paper is accepted and now you have to stand up and present it. This takes the compiled paper and produces the slide deck, along with n… | | [nnsight Remote Interpretability](/app/nnsight-remote-interpretability) | Skill | You want to see what is happening inside a neural network, but the interesting models are too big for the GPUs you have. nnsight lets you i… | | [NeMo Guardrails - Programmable Safety for LLMs](/app/nemo-guardrails) | Skill | An LLM app in front of real users can be pushed into saying things it should not, or into leaking things it should not. This adds checks ar… | | [nanoGPT - Minimalist GPT Training](/app/nanogpt) | Skill | Transformers are hard to learn from papers or from huge production codebases. This is a small, readable GPT you can go through line by line… | | [LlamaGuard - AI Content Moderation](/app/llamaguard) | Skill | If people can type anything into your app, some of it should not reach the model, and some of what the model says should not reach them. Th… | | [LangSmith - LLM Observability Platform](/app/langsmith-observability) | Skill | When an app built on a language model gives a bad answer, it is hard to see where things went wrong. This covers tracing calls, evaluating… | | [CrewAI - Multi-Agent Orchestration Framework](/app/crewai-multi-agent) | Skill | One agent working alone is not enough for the job you have. This lets you set up several agents with different roles, give them memory, and… | | [AutoGPT - Autonomous AI Agent Platform](/app/autogpt-agents) | Skill | You want an AI agent that keeps working on its own rather than answering one question and stopping. This is a platform for building those a… | | [ARA Rigor Reviewer](/app/ara-rigor-reviewer) | Skill | You have a finished research artifact and you are about to release it, but nobody has argued with it yet. This reads it critically, scores… | | [Live Research Project Manager (Live PM)](/app/ara-research-manager) | Skill | At the end of a long session, the reasoning behind what you did is still in the chat and nowhere else. This reads back over the session and… | | [Universal ARA Compiler](/app/ara-compiler) | Skill | You have research spread across PDFs, repositories, logs and loose notes, and none of it is in a form a machine can act on. This gathers th… | ## Links - [GitHub](https://github.com/Orchestra-Research) --- Canonical URL: