YAO Changsong
Community@Jensen-Yao
YAO Changsong maintains 149 skills spanning LLM training, fine-tuning, quantization, inference serving, evaluation, safety, RAG, and academic research publishing.
Agent Skills by YAO Changsong
Showing 143 vetted skills indexed across 1 GitHub repositories.
Jensen-Yaosparse-autoencoder-training
Train and analyze Sparse Autoencoders with SAELens to decompose neural network activations into interpretable features.
Jensen-Yaocnki-navigate-pages
Navigate CNKI search result pages and change sort order via browser evaluate_script.
Jensen-Yaomodel-merging
Merge fine-tuned language models with mergekit using SLERP, TIES, DARE, and task arithmetic.
Jensen-Yaodistributed-llm-pretraining-torchtitan
Configures and launches distributed LLM pretraining with torchtitan using 4D parallelism.
Jensen-Yaocnki-journal-search
Searches CNKI for journals by name, ISSN, CN, or sponsor and extracts metrics.
Jensen-Yaosystems-paper-writing
Structure systems papers for OSDI, SOSP, ASPLOS, NSDI, and EuroSys with paragraph-level blueprints.
Jensen-Yaoautoresearch
Orchestrates autonomous AI research projects using a two-loop experiment and synthesis architecture.
Jensen-Yaoimagegen
Generate and edit raster images using built-in image tools or a GPT Image CLI fallback.
Jensen-Yaoray-train
Orchestrates distributed PyTorch, TensorFlow, and HuggingFace training across multi-node Ray clusters.
Jensen-Yaounsloth
Guides LLM fine-tuning with Unsloth using LoRA, QLoRA, and GRPO techniques.
Jensen-Yaowhisper
Transcribe and translate multilingual audio using OpenAI's Whisper speech recognition models.
Jensen-Yaolong-context
Extend transformer context windows using RoPE, YaRN, ALiBi, and position interpolation techniques.
Jensen-Yaoweights-and-biases
Track ML experiments, run hyperparameter sweeps, and manage model artifacts with Weights & Biases.
Jensen-Yaoquantizing-models-bitsandbytes
Quantizes LLMs to 8-bit or 4-bit formats for reduced GPU memory usage.
Jensen-Yaodspy
Build declarative LM pipelines with automatic prompt optimization using DSPy.
Jensen-Yaopick-ui-library
Recommends one curated frontend library for a specific UI capability from a bundled list.
Jensen-Yaoopenrlhf-training
Trains large language models with distributed PPO, GRPO, RLOO, and DPO using Ray and vLLM.
Jensen-Yaonemo-guardrails
Implements programmable runtime safety rails for LLM applications using NVIDIA NeMo Guardrails.
Jensen-Yaoevaluating-cosmos-policy
Evaluates NVIDIA Cosmos Policy on LIBERO and RoboCasa robot simulation benchmarks.
Jensen-Yaollamaindex
Build RAG pipelines with document ingestion, vector indices, and query engines.
Jensen-Yaoofficecli
Create, inspect, and modify Office documents via the officecli command-line tool.
Jensen-Yaodeepspeed
Guides distributed training with DeepSpeed ZeRO optimization, pipeline parallelism, and mixed precision.
Jensen-Yaocrewai-multi-agent
Orchestrates role-based multi-agent teams with sequential or hierarchical task execution in Python.
Jensen-Yaomoe-training
Train Mixture of Experts models with DeepSpeed using sparse routing and expert parallelism.
Frequently Asked Questions About YAO Changsong
FAQPage SchemaWhat tasks can I automate using YAO Changsong's skills?▼
You can run distributed LLM pretraining (Megatron-Core, FSDP2, torchtitan), RLHF post-training (PPO, GRPO, DPO, SimPO), 2-8 bit quantization (AWQ, GPTQ, HQQ, GGUF), high-throughput serving (vLLM, SGLang, TensorRT-LLM), benchmark evaluation, RAG pipelines, and academic paper writing with figure extraction.
Who are these skills designed for?▼
ML engineers training or aligning large models, inference engineers optimizing latency and memory, interpretability researchers using TransformerLens or SAELens, robotics teams fine-tuning VLA policies (OpenVLA, OpenPI), and graduate researchers preparing conference papers, literature reviews, and thesis-defense presentations.
How do I install and run these skills in practice?▼
Each skill declares pip-installable dependencies in its frontmatter, such as torch, transformers, trl, peft, vllm, or deepspeed. Install the listed packages, then invoke the skill by name; GPU-heavy skills expect CUDA hardware, while llama.cpp and GGUF skills run on CPU or Apple Silicon.
Are these skills open-source and what do they cost?▼
Nearly all skills carry MIT licenses and wrap free open-source frameworks like DeepSpeed, FAISS, Chroma, and LLaMA-Factory; sci-figure uses AGPL-3.0. Costs arise only from your own GPU compute or optional managed services such as Pinecone, Lambda Labs, or Modal.
What prerequisites and dependencies are required?▼
Core prerequisites are PyTorch (often >=2.0) and HuggingFace Transformers; distributed skills add Ray, DeepSpeed, or Megatron-Core. Robotics skills pin exact versions (torch==2.2.0, robosuite==1.4.0), drawio-skill needs the draw.io desktop binary, and CNKI skills require an authenticated institutional login.