allenai/Olmo-core

PyTorch building blocks for the OLMo ecosystem

What it solves

Olmo-core provides the essential building blocks for modeling and training large language models (LLMs), specifically the OLMo family of models. It streamlines the process of training from scratch, continuing training (annealing), and running inference on these models.

How it works

The library provides a set of core modeling components and official training scripts. Users can launch training or annealing scripts using torchrun or the Beaker launch CLI. For inference, the project integrates with Hugging Face Transformers and vLLM for high-throughput generation, while also offering a beta autoregressive generation capability directly within the library.

Who it’s for

It is designed for AI researchers and developers who want to train, fine-tune, or deploy OLMo models, as well as those looking for a standardized way to implement LLM training pipelines.

Highlights

  • Official Training Scripts: Includes ready-to-use scripts for OLMo-2 and OLMo-3 model families (7B and 32B parameters).
  • Broad Inference Support: Compatible with Hugging Face Transformers, vLLM, and native Olmo-core generation.
  • Advanced Training Optimizations: Supports optional dependencies like FlashAttention, TransformerEngine, Liger-Kernel for memory efficiency, and torchao for float8 training.
  • MoE Support: Integrates with grouped_gemm for dropless mixture-of-experts models.

Related

  • Project
    KEV0143/Comparative-analysis-of-hourly-load-forecasting-using-PatchTST-TFT-NHiTS-and-CatBoost

    A Python research project that compares four modern time‑series models (PatchTST, TFT, N‑HiTS, CatBoost) for hour‑by‑hour electricity‑load forecasting using real Russian utility data. Includes data preprocessing, training pipelines, statistical analysis, and visualisation, and is released under Apache 2.0.

  • Project
    kyleskom/NBA-Machine-Learning-Sports-Betting

    A Python project that scrapes NBA stats and sportsbook odds, builds matchup features, trains XGBoost and TensorFlow models to predict game winners and over/under totals, and outputs expected value and Kelly‑criterion stake sizes. Includes scripts for data collection, model training, a command‑line predictor, and a simple Flask web UI.

  • Project
    ScottfreeLLC/AlphaPy

    A machine learning framework for speculators and data scientists to build predictive models for financial markets and sporting events.

  • Project
    brainflow-dev/brainflow

    A uniform SDK for obtaining, parsing, and analyzing data from biosensors such as EEG, EMG, and ECG, simplifying the development of neurointerfaces.