SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI developer platformsLLaMA-O1

LLaMA-O1

LLaMA-O1 is an open-source large reasoning model framework built on PyTorch and HuggingFace, designed to support training, inference, and evaluation. The project is developed by SimpleBerry and has released multiple pretrained and fine-tuned models and datasets, including LLaMA-O1-Supervised-1129 and others. The framework is dedicated to advancing the development of open-source large reasoning models, supporting features such as Long Chain-of-Thought (Long CoT) tokens, pretraining, supervised fine-tuning, Process Reward Model (PRM) token correction, and reinforcement learning with self-play. Some features, such as reinforcement learning, are still in training, and the inference-time reasoning enhancement framework is temporarily postponed. Users can obtain models, datasets, and an online demo (CPU only) through HuggingFace.

Open website

About this tool

LLaMA-O1 is an open-source large reasoning model framework focused on training, inference, and evaluation, built on PyTorch and HuggingFace. The project is maintained by SimpleBerry, with the goal of providing the open-source community with a complete solution for Large Reasoning Models.

Main capabilities include:

  • Long Chain-of-Thought (Long CoT) tokens: completed.
  • Pretraining dataset: completed, available on HuggingFace (OpenLongCoT-Pretrain-1202).
  • Supervised fine-tuning dataset: completed, available on HuggingFace (OpenLongCoT-SFT).
  • PRM token correction dataset: completed.
  • Reinforcement Learning With Self-Play: code completed, currently in training.
  • Inference-time reasoning enhancement framework: code completed, temporarily postponed.

Usage: Users can obtain models and datasets through HuggingFace. Specifically, these include:

  • Base model (pretrained):https://huggingface.co/SimpleBerry/LLaMA-O1-Base-1127
  • Supervised fine-tuned model:https://huggingface.co/SimpleBerry/LLaMA-O1-Supervised-1129
  • Supervised fine-tuning dataset:https://huggingface.co/datasets/SimpleBerry/OpenLongCoT-SFT
  • Pretraining dataset:https://huggingface.co/datasets/SimpleBerry/OpenLongCoT-Pretrain-1202
  • GGUF format model:https://huggingface.co/Lyte/LLaMA-O1-Supervised-1129-Q4_K_M-GGUF
  • Online demo (CPU only):https://huggingface.co/spaces/SimpleBerry/LLaMA-O1-Supervised-1129-Demo

In addition, the project also provides links to related research papers:

  • https://huggingface.co/papers/2406.07394
  • https://huggingface.co/papers/2410.02884
  • https://huggingface.co/papers/2411.18203

Applicable scenarios: LLaMA-O1 is suitable for researchers and developers to train, fine-tune, infer, and evaluate large reasoning models. It is particularly suitable for tasks that require long chain-of-thought reasoning capabilities, such as complex problem solving, mathematical reasoning, and more.

Limitations:

  • Some features (such as reinforcement learning) are still in progress and have not yet been fully released.
  • The inference-time reasoning enhancement framework is temporarily postponed.
  • The online demo only supports CPU operation, so performance may be limited.
  • Specific training details, hardware requirements, usage steps, and other information are not yet provided in the materials.

Related tools

Browse category