Training Gemma-3 for Structured Mathematical Reasoning with Tunix GRPO, LoRA Adapters, and GSM8K Rewards
In this tutorial, we build an end-to-end GRPO training workflow that teaches Gemma-3 to reason through GSM8K math problems using Tunix, JAX, LoRA, and custom reward functions. We start by preparing the environment, authenticating with Hugging Face, loading the Gemma-3 model, and wrapping GSM8K examples into a prompt format that requires both structured reasoning and a final numeric answer. We then define reward functions that assess format adherence and mathematical correctness, attach LoRA ada
This tutorial demonstrates how to train Gemma-3, an AI language model, to solve math problems by using reinforcement learning techniques including GRPO (a training method), LoRA adapters (lightweight weight adjustments), and custom reward functions. The training workflow teaches the model to structure its responses with explicit reasoning sections followed by numeric answers on GSM8K math problems. The approach matters because it shows how to improve an AI model's mathematical reasoning while keeping computational costs manageable by only updating adapter weights rather than the entire model. The tutorial requires familiarity with machine learning concepts like reinforcement learning, model training, and the ability to set up a development environment with specific libraries like JAX and Tunix.
One short email a day with the stories that matter. Free. Unsubscribe with one click.
Want to go deeper than the news? Explore live, cohort-based AI courses taught by practitioners.
Browse AI courses on Maven
Anthropic releases new Mythos and Fable models amid RSI concerns
Anthropic says the new models "show improvements in areas like coding, reasoning, and vision," as well as in communicating more clearly with users.
Also reported by The Verge, SiliconANGLE and MarkTechPost
Runway’s Praxis-1 world action model for robotics is now available
Praxis-1 builds on Runway’s work as a founding member of the Cosmos Coalition, advancing open world models for physical AI. | Source: Runway Runway AI Inc. yesterday released Praxis-1, its world action model for robotics. The model is now available to developers with open weights under the Apache 2.0 license. Runway first announced Praxis-1 in September with launch partners including Noble Machines, Standard Bots, and Ultra. Since then, these companies have been running the model on their own ha
Google releases Gemini 4 Carbon with agent ‘swarm’ feature
Google is rolling out its latest AI model, Gemini 4 Carbon, to the general public after releasing it to a select group of "cybersecurity defenders" two weeks ago. Google claims Carbon is its "strongest model" yet, and capable of competing with the very best systems from Anthropic and OpenAI. Google says Gemini 4 Carbon "pushes the frontier even further" than Argon, referring to the company's other frontier model that launched only last month. The company said Argon will likely
Gemini 4 Carbon is now available
The latest version aims to deliver improved performance while reducing the computational resources and energy required for AI operations.
Google Launches Gemini Decisions API With Gemini Decision 1 at $0.10/M Tokens
Google has launched the Gemini Decisions API, powered by a new model, Gemini Decision 1 (gemini-decision-1). The Gemini Decisions API does not generate text. It scores up to 1,000 typed questions against one shared input and returns calibrated probabilities. For developers, this targets the small branching calls inside agents: routing, policy checks, tool selection and grading. Google prices it at $0.10 per 1M input tokens, with zero output tokens billed. TL;DR Size: Context is 262,144 t
Also reported by SiliconANGLE
Introducing AlphaProof Code: a breakthrough in verifying complex software at depth
AlphaProof Code agents wrote three million lines of proofs in Lean to verify Signal’s core encryption.