Machine Learning

Run a hill-climbing study over many training runs with a budget and kill criteria

Fine-tune LLMs on Google Colab GPUs with Unsloth via a remote bridge

Train molecular property models with DeepChem featurizers and MoleculeNet data

Train region embeddings and consensus peaks from BED genomic interval files

Load PDE simulation datasets from HDF5 into PyTorch DataLoaders with downsampling

Evaluate ML models on benchmarks with clean splits, seeds and honest reporting

Turn SMILES molecules into ECFP, MACCS and pretrained model features

Train FNO and DeepONet neural operators to solve parametric PDE families fast

Train physics-informed neural networks with DeepXDE for forward and inverse PDEs

Train fast parallel reinforcement learning agents on vectorized environments

Build clinical prediction models on EHR data with the PyHealth toolkit

Load Therapeutics Data Commons datasets and benchmarks for drug ML

Fit scVI models for batch correction and multimodal single-cell integration

Explain model predictions with SHAP values, waterfall and beeswarm plots

Build neural operators that handle shocks and steep gradients in PDE solutions

Train PPO, SAC, DQN and TD3 reinforcement learning agents with Stable-Baselines3

Discover interpretable governing equations from data with PySR symbolic regression

Fine-tune LLMs with supervised, RL and LoRA training on Tinker cloud

Estimate Tinker fine-tuning costs from dataset token counts

Train graph neural networks on node, link and molecular graph tasks with PyG

Build graph neural networks for molecules and proteins with TorchDrug

Build fine-tuning datasets from production logs and frontier model outputs

Run pretrained Hugging Face transformers for text, vision and speech tasks

Fine-tune LLMs with LoRA, QLoRA and GRPO in Unsloth and export GGUF files

Run the CDF AnomalyGen pipeline to generate synthetic defect images

Author and validate Cosmos Data Factory augmentation configs for video and images

Plan and run CDF auto-labeling pipelines and their prebuilt stage containers

Run and monitor autonomous vehicle dataset generation Airflow workflows

Curate video and image datasets by splitting, deduplicating and captioning them

Generate synthetic PCB inspection images with defect textures on Kubernetes

Generate anomaly event videos from images through the CDF orchestration DAG

Augment clothing and person attributes in images for re-identification data

Write custom CDF Orchestration DAGs that chain selected pipeline stages

Create synthetic PCB inspection images with golden boards and solder defects

Create Cosmos video variants and auto-label source and generated videos

Write GPU data-loading code with the DALI imperative dynamic mode in Python

Generate synthetic datasets and build data generation pipelines from a spec

Run FoundationPose object pose estimation on BOP datasets and score results

Install FoundationPose and build FoundationStereo TensorRT depth engines

Grade workflow episodes with a vision model to label successful demonstrations

Convert workflow HDF5 recordings into LeRobot datasets for training and viewing

Expand demonstrations into synthetic variants with action jitter on segments

Record robot demonstrations through keyboard, leader, VR or bus teleop input

Run the workflow pipeline from recording to validated policy checkpoints

Fine-tune GR00T or openpi robot policies on compatible LeRobot data

Train, evaluate and export workflow policies with RSL-RL or RLinf

Choose an LLM serving stack and memory flags for Jetson devices

Benchmark LLM and VLM serving on Jetson across vLLM, llama.cpp and Ollama

Stand up vLLM or SGLang model serving on Jetson boards