Machine Learning

Run production ML work with data contracts, reproducible training, evals and rollback

Build reproducible PyTorch training loops, model architectures and data loaders

Run PyTorch training across multiple GPUs with minimal code changes

Fine-tune LLMs from YAML configs with LoRA, DPO and GRPO using Axolotl

Classify images zero-shot and search images by text with CLIP

Speed up long-sequence transformer training and inference with FlashAttention

Search, download and upload Hugging Face models and datasets with hf

Tokenize text fast with BPE or WordPiece and train custom vocabularies

Run GGUF models locally with llama.cpp and find models on the Hugging Face Hub

Run LLaVA vision-language chat for visual questions and image captions

Curate LLM training data by deduplicating, filtering and redacting PII

Analyze and ablate refusal directions in open-weight LLMs with diff-in-means

Fine-tune large language models with LoRA on limited GPU memory

Train large models with PyTorch fully sharded data parallelism

Write clean PyTorch training loops with built-in distributed support

Train sparse autoencoders to interpret features inside language models

Segment objects in images zero-shot using points, boxes or masks with SAM

Serve LLMs at high throughput with vLLM and an OpenAI-compatible API

Align language models with reference-free SimPO preference training

Run reinforcement learning post-training for LLMs with Megatron and SGLang

Pretrain large language models at scale with PyTorch 4D parallelism

Fine-tune language models with SFT, DPO, GRPO and reward modeling in TRL

Fine-tune models with LoRA or QLoRA faster and with less GPU memory using Unsloth

Log ML experiments, run sweeps and manage models with Weights and Biases

Build a small test slice of a Hugging Face model by fetching only needed tensors

Train RuView pose, embedding and domain models on WiFi sensing data

Train SONA and MicroLoRA neural patterns from completed successful tasks

Train LSTM, Transformer and N-BEATS models on market data with confidence intervals

Place hierarchical data in a Poincare ball using depth for radius

Automate Bigml tasks with the Rube MCP toolkit from Composio

Automate Datarobot tasks with the Rube MCP toolkit from Composio

Automate Kaggle tasks with the Rube MCP toolkit from Composio

Read license plates from images with Plate Recognizer through Composio Rube MCP

Run Replicate AI model predictions, upload files and inspect model schemas

Train and run Uni-Mol molecular property prediction models from a CLI

Turn bad agent cases into DPO preference pairs for LoRA fine-tuning

Pick the right post-training route for an agent model, from SFT to RL

Design reward functions for RL training and catch reward hacking early

Classify, forecast and detect anomalies in time series with specialized algorithms

Refine hypotheses through repeatable experiments and held-out candidate comparisons

Build molecular property models with DeepChem for toxicity, solubility and ADMET

Find and evaluate scientific datasets and models in the Hugging Science catalog

Turn small molecules into fingerprints, descriptors and embeddings for QSAR models

Build and train reinforcement learning environments with PufferLib and PuffeRL

Organize PyTorch training code into Lightning modules, Trainers and distributed runs

Build scikit-learn pipelines for classification, regression, clustering and tuning

Build censoring-aware survival models with scikit-survival and leakage-safe checks

Explain machine learning predictions with SHAP values and feature attribution plots

Train reinforcement learning policies with Stable Baselines3 PPO, SAC and DQN