Tina is a research project focused on small reasoning models, and its paper "Tina: Tiny Reasoning Models via LoRA" has been accepted by ICLR 2026. The project's core goal is to answer a key question: how can reasoning abilities be efficiently injected into language models through reinforcement learning in a cost-effective way?
Specifically, Tina explores the use of low-rank adaptation (LoRA) during reinforcement learning to enhance the reasoning abilities of small language models. Experimental results show that, on the same base model, the Tina models can match, and in some cases even exceed, the performance of SOTA baseline models trained with full parameters. Notably, the best Tina model achieved a performance improvement of more than 20% on the AIME24 benchmark, with a Pass@1 accuracy of 43.33%. In terms of cost, reproducing the best Tina checkpoint requires only about $9, while the total cost of reproducing all experiments from scratch is about $526.
The project code is built on HuggingFace's open-r1 project and uses the GRPO algorithm for LoRA reinforcement learning training. The repository provides the main environment variable scripts, recipes for each experiment, core configuration files, and dataset constants. Users can modify the relevant paths and HuggingFace Hub ID according to their own setup. The training and evaluation workflow is executed through scripts, including LoRA-based RL with GRPO training, re-evaluation of baseline models, and evaluation of post-trained models.
The project's applicable scenarios are mainly research environments, suitable for researchers and developers interested in reinforcement learning, parameter-efficient fine-tuning, and improving the reasoning abilities of small language models. Because it involves specific training and evaluation steps, using it requires a certain foundation in machine learning and deep learning, and the environment and dependencies must be configured according to the documentation. The information currently provided by the project mainly consists of the code repository and the paper. For more details on model specifics, dataset usage, and experimental settings, it is recommended to refer to the original paper and the configuration files in the repository.
