Machine Learning

Build graph neural networks for node, link and graph tasks in PyTorch Geometric

Load Hugging Face models, run pipelines and fine-tune with the Transformers Trainer

Reduce dimensions and build 2D or 3D embeddings with UMAP for clustering

Gate fine-tuned checkpoints with drift budgets and forgetting checks before release

Format, validate and document datasets for fine-tuning with chat templates

Build the eval harness, golden sets and graders that gate every fine-tuning run

Decide whether to fine-tune and choose SFT, DPO, GRPO or continued pretraining

Train models on verifiable tasks with GRPO and RLVR reward design

Configure LoRA and QLoRA fine-tuning with rank, alpha and target module choices

Build MLOps pipelines from data prep through training, validation and deployment

Align models with DPO, ORPO, KTO or SimPO using preference pairs

Export fine-tuned models as merged safetensors, LoRA, GGUF or FP8 for deployment

Design top-K recommendation and feed pipelines in composable ranking stages

Install PyTorch, Unsloth, TRL and vLLM on NVIDIA DGX Spark aarch64

Manage unified memory headroom and thermals on DGX Spark during long training runs

Diagnose the ten common failure modes of ML training on DGX Spark

Turn eval traces and production logs into SFT examples and preference pairs

Fine-tune vision-language models on image and text data with LoRA

Train predictive models that score users daily for churn, activity and conversion

Fine-tune a model with GRPO on managed GPUs from a task and a dataset

Download, upload and manage Hugging Face models, datasets and repos with the hf CLI

Create Hugging Face datasets with configs, streaming updates and SQL queries

Add evaluation results to Hugging Face model cards from READMEs and benchmark APIs

Run Python scripts and Docker workloads on Hugging Face Jobs with GPU hardware

Fine-tune language models with SFT, DPO and GRPO using TRL on cloud GPUs

Build reusable scripts that fetch and process data from the Hugging Face API

Log training metrics and visualize ML experiments with Trackio dashboards

Find poisoned data and backdoored models using ART, Cleanlab and safetensors checks

Catch model stealing from inference APIs by tracking query patterns and rate-limiting

Forecast time series zero-shot with Google's TimesFM foundation model

Train reinforcement learning policies with fast parallel simulation in PufferLib

Build computer vision pipelines for detection, segmentation and video with YOLO

Deploy ML models with MLOps, feature stores, monitoring and RAG systems

Train reinforcement learning models with Stable-Baselines3 PPO, SAC and DQN

Train GPU-accelerated classification, regression and clustering models with cuML

Build object detection and segmentation systems with YOLO, SAM and ONNX

Deploy ML models with MLOps pipelines, drift monitoring and RAG systems

Apply the Spinning Up deep RL guide to policy gradients, PPO, SAC and debugging

Write BigQuery SQL for forecasting, anomaly detection, vector search and text classification

Deploy open or custom models from Model Garden to Agent Platform endpoints

Create, update and delete serving endpoints for models on Agent Platform

Upload, version and manage machine learning models in the Agent Platform registry

Fine-tune open and Gemini models on Agent Platform infrastructure

List, inspect and cancel model tuning jobs on Agent Platform

Build a book-to-SFT pipeline with ePub extraction, segmentation and LoRA training

Fine-tune LLMs with LoRA and QLoRA, including dataset prep and tuning

Train, evaluate and ship machine learning models with PyTorch and scikit-learn

Turn an experiment plan into code, deploy it to a GPU and collect first results

Queue multi-seed ML experiments over SSH with OOM-aware retries and cleanup