Fine-tuning, addestramento e dati

Oumi: Open-Source Foundation Model Training, Evaluation, Inference, and Deployment

oumi-ai/oumi

Oumi is a Python-based platform for preparing data, training and fine-tuning open-weight foundation models, evaluating results, running inference, and deploying models. It provides configuration recipes and a consistent CLI for local, cluster, and cloud workflows.

Apri il repository ufficiale ↗ Pagina iniziale del progetto ↗
★ 9,4KStelle
⑂ 784Fork
31Problemi aperti
PythonLingua
Apache-2.0Licenza
Q@project.QualityScorePunteggio editoriale

Panoramica

The project covers the foundation-model lifecycle from data synthesis and curation through training, evaluation, inference, remote job execution, and deployment. Supported techniques described in the README include full fine-tuning, SFT, LoRA, QLoRA, DPO, KTO, GRPO, vision DPO, model distillation, and AWQ quantization. Oumi supports text and multimodal workflows, distributed training through FSDP, DeepSpeed, and DDP, and inference through engines including vLLM and SGLang. The repository metadata reports 9,369 stars, 784 forks, 31 open issues, and an active, non-archived main branch.

Funzionalità principali

  • Training and fine-tuning for models ranging from 10M to 405B parameters, as stated in the README
  • Full fine-tuning, SFT, LoRA, QLoRA, DPO, KTO, GRPO, and vision-oriented workflows
  • Text and multimodal model support
  • Ready-to-use YAML recipes for training, evaluation, and inference
  • Data synthesis and curation with LLM judges
  • Distributed training with FSDP, DeepSpeed, and DDP
  • Inference integrations including vLLM and SGLang
  • Remote jobs across AWS, Azure, GCP, and Lambda
  • Commercial API integrations including OpenAI, Anthropic, Vertex AI, Together, and Parasail
  • Dedicated endpoint deployment through the oumi deploy CLI
  • MCP server integration for Claude and Cursor
  • Notebook tutorials and Google Colab examples

Requisiti, installazione e avvio rapido

Recommended package installation: `uv pip install oumi`. For GPU support: `uv pip install 'oumi[gpu]'`. The development version can be installed with `uv pip install git+https://github.com/oumi-ai/oumi.git`; the README says pip may be used instead of uv. A container image is available with `docker pull ghcr.io/oumi-ai/oumi:latest`. The README also provides an experimental isolated installer: `curl -LsSf https://oumi.ai/install.sh | bash`. Review scripts before execution and consult https://oumi.ai/docs/en/latest/get_started/installation.html for advanced options.

Utilizzo

Quick start with the included recipes: train using `oumi train -c configs/recipes/smollm/sft/135m/quickstart_train.yaml`; evaluate using `oumi evaluate -c configs/recipes/smollm/evaluation/135m/quickstart_eval.yaml`; and start interactive inference using `oumi infer -c configs/recipes/smollm/inference/135m_infer.yaml --interactive`. Remote jobs use `oumi launch up -c <config>`, with the target cloud selectable through `--resources.cloud`. Tutorial notebooks cover the core workflow, LoRA fine-tuning, distillation, evaluation, remote jobs, and LLM-as-a-Judge curation.

Compatibilità del modello e casi d'uso

The supplied record explicitly mentions Qwen, Gemma, Llama, DeepSeek, Phi, Falcon, OLMo, SmolLM, gpt-oss, and vision-language models. Recipe examples include Qwen3 and Qwen vision variants, DeepSeek R1 and distilled variants, Llama 3.x and Llama 4, Falcon-H1, Falcon-E, Gemma 3, and OLMo 3. News entries report support for Gemma 4 and Qwen3.5. The README says compatibility is broader than its example recipes, but an exhaustive model list is not included in the supplied excerpt.

Note su licenza e rischi

Repository metadata identifies the license as Apache-2.0. License URL: https://api.github.com/licenses/apache-2.0. Dependencies, model weights, datasets, hosted APIs, and provider services may have separate terms; their license details are not stated in the repository metadata.

Rilascio e manutenzione

The README reports Oumi v0.8 with `oumi deploy`, an `oumi-mcp` server, batch API support across Anthropic, Fireworks, and Together, plus Transformers v5, TRL, and vLLM upgrades. It also reports Gemma 4 and Qwen3.5 support. Earlier listed releases added Python 3.13 support, data analysis, synthesis, hyperparameter tuning, DeepSpeed, vision DPO, AWQ quantization, adaptive inference, and GRPO. Release page: https://github.com/oumi-ai/oumi/releases/tag/v0.8.

Langfuse

langfuse/langfuse

★ 32,4KTypeScript