Model Training & Fine-tuning Agent Skills

Navigate the PhysicsNeMo repository by discovering model families, datapipes, and examples through live file search, without writing training code.

Train, evaluate, export, and run inference for CenterPose models used in 6-DoF object pose estimation with keypoint regression.

Run evaluation tasks against a NeMo Platform server using the Evaluator plugin CLI and Python SDK.

Lists TAO Skill Bank capabilities, models, and AutoML support by running packaged scripts.

Fine-tune a TAO model with standard supervised training, evaluation, and export, with AutoML bypass and platform-specific credential intake.

Install Holoscan SDK v4.3+ via Conda in a CUDA 13 environment, including Python bindings and C++ development headers.

Fine-tune NV-Segment-CT VISTA3D on CT NIfTI labels for smoke testing or dataset adaptation, wrapping the upstream MONAI bundle entrypoint.

Trains, evaluates, exports, and runs inference for NVPanoptix3D models that perform panoptic 3D scene reconstruction from posed RGB images, producing 3D panoptic segmentation with occupancy completion.

Train, evaluate, export, prune, and run inference for PointPillars 3D object detection models from LiDAR point clouds using NVIDIA TAO.

Configure fault tolerance, straggler detection, preemption, in-process restart, and re-run state machine for Megatron Bridge training jobs.

Run TAO SDK jobs as Docker containers on a local or remote Docker daemon with NVIDIA GPU support, including preflight checks and credential handling.

Extract false-positive and false-negative gaps from VLM binary-classification-question predictions by comparing model responses against ground truth, producing a structured JSONL file and summary report for downstream root-cause analysis.

Trains, evaluates, exports, quantizes, and runs inference for a Grounding DINO model that detects objects described by text prompts without a fixed class vocabulary.

Recommends an inference runtime and memory-related launch flags for LLM/VLM workloads on NVIDIA Jetson devices, based on a live memory audit snapshot.

Performs deep root cause analysis on NVIDIA TAO Visual ChangeNet classification experiments, using image-evidence-driven investigation to diagnose model failures and produce actionable reports.

Fine-tune, evaluate, run inference, and export Cosmos-Embed1 video-text embedding models for tasks like text-to-video retrieval and semantic deduplication.

Trains, evaluates, and runs inference for Mask Auto-Label (MAL) weakly-supervised segmentation models using ViT-MAE backbones with minimal point or box annotations.

Adds autotuning to CuTile kernels using the exhaustive_search API with a tune-once/cache/direct-launch pattern, covering occupancy-only and complex tile-size search spaces.

Plan, debug, tune, evaluate, export, or deploy public Nemotron embedding and reranking retrieval recipes using the current checkout.

Fine-tune Cosmos Reason video QA models using supervised fine-tuning with FSDP parallelism, including dataset preparation, spec construction, and AutoML support.

Integrate a HuggingFace computer vision model into the NVIDIA TAO Toolkit ecosystem, covering the full pipeline from prerequisites to container testing.

Reduce per-token latency on Jetson vLLM servers by appending speculative decoding configuration, with guidance on when to enable and how to benchmark the improvement.

Train, evaluate, export, and run inference for monocular depth estimation models using Metric Depth Anything v2 or Relative Depth Anything architectures via the TAO toolkit.

Trains, evaluates, exports, and runs inference on FoundationStereo models for stereo depth estimation and 3D reconstruction from stereo image pairs.

Train, evaluate, export, and run inference for Masked Auto-Encoder (MAE) models for self-supervised pretraining and fine-tuning of visual representations.

Train, evaluate, export, and run inference on TAO action-recognition models for classifying temporal actions in video clips using RGB, optical flow, or joint input.

Trains, evaluates, exports, and runs inference for Siamese-network-based optical inspection models to detect manufacturing defects and quality issues in image pairs.

Validate and use CUDA graph capture in Megatron Bridge, including local full-iteration graphs and Transformer Engine scoped graphs for attention, MLP, and MoE modules.

Finetunes the NV-Generate-CTMR MR-brain diffusion UNet from user-supplied NIfTI training volumes using a wrapper that stages configs and delegates to upstream scripts.

Trains, evaluates, exports, quantizes, and runs inference for a Mask Grounding DINO model for open-set instance segmentation guided by text prompts.

Import object detection models from HuggingFace or NVIDIA NGC into a DeepStream pipeline with automated ONNX download, TensorRT engine build, custom parser, multi-stream benchmark, and PDF report generation.

Fine-tune HuggingFace CV, VLM, or LLM models on local NVIDIA GPUs using an NGC PyTorch container, with support for full or LoRA training, dataset handling, and optional model push to the Hub.

Train, evaluate, distill, quantize, export, and run inference for PyTorch-based TAO image classification models with support for multiple backbones.

Assesses NumPy code for cuPyNumeric migration readiness by analyzing source code against an API support manifest and GPU-scaling idioms, producing a structured verdict with per-finding reasoning.

Guides implementation of new model architectures in NeMo AutoModel through five phases: discovery, implementation, registration, validation, and testing.

Enables Megatron Fully Sharded Data Parallel in Megatron-Bridge with configuration overrides, code anchors, pitfalls, and verification steps.