Model Training & Fine-tuning Agent Skills

Optimizes MoE expert-parallel communication overlap in Megatron Bridge, covering dispatch/combine overlap, flex dispatcher backends, and expert wgrad scheduling.

Generates Python code using BigQuery DataFrames (BigFrames), the pandas/scikit-learn-style API over BigQuery, for dataframe and ML workflows.

Create Earth2Studio prognostic model wrappers that time-step weather forecasts forward, with triple-inheritance classes, tests, and documentation.

Quantize large language models to 4-bit using activation-aware weight quantization, achieving ~3x speedup with minimal accuracy loss for deployment on limited GPU memory.

Convert and quantize models to GGUF format for efficient CPU/GPU inference with llama.cpp, supporting 2-8 bit quantization and Apple Silicon acceleration.

Fine-tune and serve Physical Intelligence OpenPI models (pi0, pi0-fast, pi0.5) using JAX or PyTorch backends for robot policy inference across ALOHA, DROID, and LIBERO environments.

Train large-scale MoE models with FP8/INT4 low-precision RL, speculative decoding, and train-inference alignment using the miles framework.

Run ML training and batch jobs across multiple clouds with automatic cost optimization, spot instance recovery, and unified orchestration.

Train reinforcement learning models using torchforge, Meta's PyTorch-native RL library for scalable, algorithm-focused experimentation with GRPO, DAPO, and custom loss functions.

Provides expert guidance for fine-tuning LLMs with LLaMA-Factory, covering WebUI no-code, 100+ models, 2/3/4/5/6/8-bit QLoRA, and multimodal support.

Train and use SentencePiece tokenizers for multilingual NLP, supporting BPE and Unigram algorithms with raw Unicode text.

Pretrains large language models from scratch using PyTorch-native distributed training with 4D parallelism (FSDP2, TP, PP, CP) and Float8 support on H100 GPUs.

Scales machine learning training from single GPU to multi-node clusters with minimal code changes. Supports PyTorch, TensorFlow, and HuggingFace with built-in hyperparameter tuning, fault tolerance, and elastic scaling.

Evaluates LLMs across 60+ academic benchmarks (MMLU, HumanEval, GSM8K, TruthfulQA, HellaSwag) using standardized prompts and metrics. Supports HuggingFace, vLLM, and API backends.

Organizes PyTorch code with a Trainer class, automatic distributed training (DDP/FSDP/DeepSpeed), callbacks, and minimal boilerplate. Scales from laptop to supercomputer with the same code.

End-to-end Kaggle competition workflow for scored submissions, covering code competitions, validation, metrics, public notebook/discussion intel, tabular/text/image modeling, tuning, ensembling, multi-notebook architectures, Kaggle GPU offload, and hidden-test debugging.

Generates and analyzes proteins using ESM3 and ESM C language models, covering sequence generation, structure prediction, inverse folding, embeddings, and function conditioning with local or cloud-based Forge API inference.

Provides scikit-learn compatible algorithms for time series machine learning, including classification, regression, clustering, forecasting, anomaly detection, segmentation, and similarity search.

Classifies situations into 256 binary states and maps each to a prescribed action, reporting the pattern, decimal, name, range, and action to execute.

Generates and analyzes protein sequences and structures using ESM3, ESMC, and ESMFold2, with support for local and cloud inference.

Provides guidance for using RDKit to read and write molecular structures, calculate descriptors, generate fingerprints, perform substructure searches, and handle chemical reactions.

Estimates memory requirements for running Hugging Face models, including optional KV cache, using HTTP range requests without downloading weights.

Implements machine learning models, embeddings, and AI-powered features with ethical considerations, including model selection, integration, and monitoring.

Fine-tune large language models by training less than 1% of parameters using LoRA, QLoRA, and 25+ adapter methods, enabling efficient adaptation on limited GPU memory.

Provides domain knowledge on the Causal Dynamics Architecture (CDA), an alternative AI computing architecture based on causal graphs and Hamiltonian dynamics, with references for deep dives.

Predicts chromatin state across 919 tracks (histone marks, DNase, TF binding) for DNA sequences via the hosted Genomic Intelligence API, producing a report and JSON results.

Launches a model evaluation batch with parameter collection, pre-flight checks, execution, and post-run analysis for interactive or foreground runs.

Scales AI and Python applications across clusters with distributed computing primitives for ML workloads.

Guides the implementation of ethical AI principles, including fairness auditing, bias mitigation, explainability, accountability, and privacy protection in machine learning systems.

Implements and applies algorithms concepts, designs solutions using algorithmic principles, and optimizes performance for algorithm implementations.

Designs and implements RAG pipelines, covering document chunking, embedding strategies, hybrid search, answer synthesis with source attribution, and evaluation using RAGAS metrics.

Provides guidance on using the Aeon library for time series forecasting, covering model selection, implementation, and evaluation.

Orchestrates AI/ML workflows for building LLM applications, RAG systems, AI agents, and ML pipelines, covering design, integration, observability, and security.

Formalizes narrative technical documents into structured mathematical problem specifications with variables, constraints, objectives, and uncertainty, citing evidence and flagging missing information.

Quantize LLMs to 8/4/3/2/1-bit precision without calibration data, using multiple backends and HuggingFace/vLLM integration.

Validates the Emotional Attitude Score (EAS) metric by measuring its consistency with human judgment on word-level sentiment polarity, using the AmbGIMT dataset and pairwise score comparisons.