EVENT DOSSIER
Mitra-v2 is a new tabular base model trained on synthetic data, with 77 million parameters. The model uses a small 2D Transformer architecture, enabling support for longer contexts and a larger feature space. In the TabArena and TALENT benchmarks, its performance surpassed TabPFN-3 and it performed leading in multi-class classification tasks. Although pre-training involves only up to ten classes, Mitra-v2 achieved comparable frontier performance with a parameter size that is only 5% of TabFM’s 1.6B. The research team has released the model weights, inference, and fine-tuning code, and made it open source under the Apache-2.0 license.
arXiv Mitra-v2 开源发布基于合成数据的表格基础模型
Mitra-v2 是一款基于合成数据训练的新型表格基础模型，拥有 7700 万参数，在 TabArena 和 TALENT 基准测试中达到行业顶尖水平。该模型采用小型 2D Transformer 架构，支持更长上下文和更大特征空间，其性能超越 TabPFN-3，并在分类任务上领先于拥有超过十类的场景（尽管预训练仅涉及最多十类）。Mitra-v2 在参数规模仅为 TabFM 1.6B 的 5% 时，实现了与其相当的 frontier 性能。研究团队已发布模型权重、推理及微调代码，并采用 Apache-2.0 许可协议开源。