Machine Learning

Quantize LLMs to 8-bit or 4-bit with bitsandbytes to cut memory use

Run distributed training and hyperparameter sweeps across clusters with Ray Train

Use RWKV RNN-style models for linear-time inference with unlimited context

Generate sentence, text and image embeddings for search with Sentence Transformers

Train language-independent SentencePiece tokenizers for multilingual and CJK text

Serve LLMs with vLLM for high-throughput, OpenAI-compatible inference APIs

Serve LLMs with SGLang for structured JSON, regex output and prefix-cached inference

Align LLMs with SimPO, a reference-free alternative to DPO that trains faster

Post-train LLMs with reinforcement learning using slime on Megatron and SGLang

Train and analyze sparse autoencoders to find interpretable features in model activations

Speed up LLM inference with speculative decoding, Medusa heads and lookahead decoding

Visualize training metrics, histograms, model graphs and profiles with TensorBoard

Optimize LLM inference on NVIDIA GPUs with TensorRT-LLM for high throughput

Build PyTorch-native RL training with torchforge that separates infra from algorithms

Train very large LLMs with Megatron-Core tensor, pipeline and expert parallelism

Inspect transformer internals with TransformerLens hooks, activation caching and patching

Train LLMs with RLHF, GRPO and PPO at scale using the verl framework

Log ML runs, run hyperparameter sweeps and manage models with Weights and Biases

Fine-tune LLMs with LoRA or QLoRA adapters and JSONL datasets using Hugging Face PEFT

Set up MLflow tracking, Airflow training DAGs, feature stores and model retraining

Pick the right fine-tuning framework for MiniCPM5 SFT, LoRA or DPO training

Train a LoRA adapter for MiniCPM5-1B and convert it to GGUF for llama.cpp

Fine-tune MiniCPM5 with LLaMA-Factory YAML configs, SFT, DPO and WebUI

Fine-tune MiniCPM5 with ms-swift and Megatron-SWIFT using the right template

Fine-tune MiniCPM5 with bare TRL and PEFT, including assistant-only loss

Run low-VRAM LoRA and QLoRA fine-tuning of MiniCPM5 on one consumer GPU with Unsloth

Fine-tune MiniCPM5-1B with xtuner using config-file-driven training

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

Estimate memory needed to load Safetensors or GGUF model weights from the Hub

Find the best model for a task and compare models by benchmark scores

Run model evaluations locally with inspect-ai and lighteval on your own GPUs

Fine-tune language and vision models with TRL or Unsloth on cloud GPUs

Run GGUF models locally with llama.cpp and pick quantizations for your hardware

Build and publish a Gradio demo on Hugging Face Spaces for a LoRA adapter

Log training metrics, alerts and dashboards for ML experiments with Trackio

Train object detection, image classification and segmentation models on cloud GPUs

Fine-tune sentence embedding and reranker models for retrieval and similarity

Run Hugging Face models in JavaScript or TypeScript for text, vision and audio tasks

Post-train LLMs with SFT, DPO, GRPO, KTO and reward models using the TRL library

Run bounded experiment loops that keep only the changes that beat a benchmark

Find implementable ML training recipes from papers, datasets and code

Build machine learning models with scikit-learn, PyTorch and TensorFlow

Upload evaluation result folders to Hugging Face dataset repos with sparse checkout

Apply Factory and Registry patterns when adding new registrable ML components

Learn techniques from winning Kaggle solutions across NLP, vision and tabular data

Choose and implement AI and ML features in .NET with ML.NET, MEAI or ONNX

Solve multi-objective engineering problems with NSGA-II and Pareto fronts

Build classification, clustering and ML pipelines with scikit-learn in Python

Forecast any univariate time series zero-shot with Google TimesFM and get intervals