|
|
--- |
|
|
base_model: |
|
|
- sapientinc/HRM-checkpoint-ARC-2 |
|
|
- sapientinc/HRM-checkpoint-sudoku-extreme |
|
|
- sapientinc/HRM-checkpoint-maze-30x30-hard |
|
|
- google/flan-t5-small |
|
|
--- |
|
|
|
|
|
HRM-LLM: A truly decentralized, human-like reasoning model built by the community |
|
|
|
|
|
HRM-LLM is a community-driven large language model powered by the Hierarchical Reasoning Model (HRM) architecture. It aims to be truly decentralized: anyone can train, contribute, and scale it forward from anywhere. HRM-LLM is designed to think and work like a human—iterating, refining, and allocating compute adaptively—so it learns efficiently and generalizes across tasks. |
|
|
|
|
|
Why HRM-LLM? |
|
|
- Human-like reasoning core: HRM brings hierarchical representations and adaptive computation to mimic iterative human thinking and planning. |
|
|
- Adaptive Computation Time (ACT): The model dynamically decides how much “thought” to spend per token—more for hard tokens, less for easy ones. |
|
|
- Decentralized and scalable: Anyone can hop in, train a few steps, and push a unified checkpoint to the Hub. Every contribution compounds. |
|
|
- Simple, hackable stack: PyTorch + Transformers + Datasets. Easy to extend, easy to improve. |
|
|
- Community-aligned progress: Transparent training, open checkpoints, and community governance. |
|
|
|
|
|
What this model aims to do |
|
|
- Break down complex problems into stages, reason across them, and refine answers over multiple internal steps. |
|
|
- Learn efficient patterns via ACT, saving compute where possible and spending it where it matters most. |
|
|
- Become a robust, general-purpose assistant shaped by its global community of contributors. |
|
|
|
|
|
How you can help |
|
|
- Train a few steps in Colab (or locally) and push your contribution. |
|
|
- Experiment with hyperparameters, tokenizers, datasets, or new HRM blocks. |
|
|
- Share insights and logs to improve the next iteration. |
|
|
|
|
|
License |
|
|
- This project is licensed under Apache-2.0. You’re free to use, modify, and distribute—with attribution and notice. |
|
|
|
|
|
Jump in and train |
|
|
- Colab (1-click): https://colab.research.google.com/drive/1xZNYC-yhwdJxzbpwRekE_rDjTki5CvEv?usp=sharing |
|
|
|
|
|
Quick start: contribute training from your environment |
|
|
Run this to join training and push your contribution to the shared checkpoint. |
|
|
|
|
|
That’s it—share the Colab link, invite contributors, and let the community grow HRM-LLM together. |
