Mitra-v2 Regressor

Mitra-v2 Classifier is a tabular foundation model from AutoGluon and Amazon for classification tasks. It is pretrained entirely on synthetic datasets sampled from random classifier priors, including a Hybrid SCM prior, rather than on real-world datasets. The model uses a 12-layer 2D Transformer with attention across both rows and columns and incorporates in-context learning. Its recommended recipe fine-tunes 50 steps on a user's table and bags eight model copies for prediction; it can be run through the mitra-finetune package or used as a Mitra model in AutoGluon Tabular. The repository describes it as the second generation of Mitra, with a longer context, more features, and an improved optimizer than Mitra-v1. The weights are distributed on Hugging Face under the Apache-2.0 license and require a CUDA GPU for the documented fine-tuning and eight-fold bagging recipe.

mitra-finetune is a Python package for fine-tuning and running inference with second-generation Mitra v2 tabular foundation-model checkpoints. It supports classification and regression through the `MitraFinetune` API, including `fit`, `predict`, `predict_proba`, and distributional regression outputs. Each fit performs a 50-step full fine-tuning run through AutoGluon's eight-fold bagging wrapper. The package uses capped, optionally class-balanced in-context support for large tables, narrows wide tables with top-K feature selection or truncated-SVD projection, and uses hierarchical label decomposition for classification problems exceeding the checkpoint's native 10-class head. Regression is represented as classification over 1,000 target bins, with predictions returned as point estimates or predictive distributions. The package requires Python 3.11–3.13, a CUDA GPU, AutoGluon with the Mitra extra, and the TabArena execution wrapper. It uses checkpoints hosted on Hugging Face, including `autogluon/mitra-classifier-2` and `autogluon/mitra-regressor-2`; FlashAttention is optional. The project is distributed from its Hugging Face repository and can be installed from a cloned checkout.

AutoGluon is an open-source Python machine-learning library for automated modeling of tabular data, time series, and other data types. It trains end-to-end predictive models by selecting and combining suitable algorithms, including classic machine-learning models, ensembles, and foundation models; its tabular API can fit a predictor from a training file and generate test-set predictions with a few lines of code. The library supports Python 3.10–3.13 on Linux, macOS, and Windows, with optional GPU support, and is licensed under Apache 2.0.

MiniMind is an open-source tiny large-language-model project and end-to-end training tutorial developed by Jingyao Gong. Its main dense model is approximately 64M parameters, and the repository provides the model architecture, tokenizer, datasets, inference code, and training pipeline. The pipeline covers pretraining, supervised fine-tuning, hand-written LoRA, DPO, PPO, GRPO, CISPO, model distillation, tool calling, adaptive thinking, and agentic reinforcement learning. Core algorithms are implemented directly with native PyTorch rather than relying on high-level abstractions from third-party training libraries. Its agentic reinforcement-learning path performs multi-turn rollouts, executes generated tool calls, appends tool observations to the context, calculates trajectory-level rewards, and updates the policy; rollout can use local PyTorch generation or an SGLang server. MiniMind supports dense and mixture-of-experts variants, single- and multi-GPU training, YaRN-based RoPE length extrapolation, Transformers-format models, and inference through llama.cpp, vLLM, or Ollama. The repository also includes a Streamlit chat interface and a lightweight OpenAI-compatible API server with tool-call and reasoning fields. It is released under the Apache License 2.0.

Needle 2 is an open-source, 45-million-parameter foundation model for tool calling, device use, and structured extraction on small devices. Its weights are compressed into a single 14 MB binary and integrated with its inference engine; the repository reports that a full session uses about 28 MB of RAM and performs inference without network access after setup. The model converts text and tool descriptions into schema-constrained JSON calls using a byte-level grammar, attaches a learned confidence score for threshold-based escalation, and can retrieve the top five tools from a larger catalogue. It uses a 256-token sliding window with tools pinned as key-value sinks to bound memory, and its Simple Attention Network combines a Hadamard MLP, grouped-query attention, engram key-value memory, and multi-lane hyper-connections. The Python package supports inference, LoRA fine-tuning, and checkpoint export; the inference engine is fetched from Hugging Face and cached, while the training stack is an optional installation.