rony/gemma3-270m-typed-decisions

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.3BQuant:BF16Context Size:32kPublished:Oct 12, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

The rony/gemma3-270m-typed-decisions model is a 270 million parameter Gemma 3 architecture, fully fine-tuned by rony as a probabilistic decision model. It is specifically optimized for the LocalLLaMA/typed-decisions dataset, excelling at structured decision-making tasks by predicting answer distributions. This model achieves an accuracy of 0.761 and a KL divergence of 0.1319 on unseen test cases, demonstrating its capability in providing probabilistic outputs for various question types.

Loading preview...

Overview

The rony/gemma3-270m-typed-decisions model is a specialized 270 million parameter Gemma 3 variant, fine-tuned by rony to function as a probabilistic decision model. It was trained on the LocalLLaMA/typed-decisions dataset, focusing on generating probability distributions for structured decision-making scenarios rather than single-token answers.

Key Capabilities

  • Probabilistic Decision Making: Outputs a probability distribution over possible answers for a given question, indicating confidence levels.
  • Structured Input Processing: Designed to interpret JSON-formatted case data and question instructions to inform decisions.
  • Workflow Specialization: Evaluated across diverse workflows including agent trace observability, customer service, invoice processing, and security incidents.
  • Efficient Inference: Achieves a latency of 68 ms per case on CUDA, making it suitable for applications requiring quick probabilistic assessments.

Performance Highlights

On the LocalLLaMA/typed-decisions all/test split (400 cases, 2000 decisions), the model demonstrates:

  • Accuracy: 0.761
  • KL from Gold: 0.1319
  • Brier Score: 0.0716
  • ECE: 0.1480 (0.1146 pooled over all decisions)

This performance significantly surpasses baselines, with a prior accuracy of 0.478 and zero-shot Gemma 3 270M accuracy of 0.279.

Good For

  • Applications requiring probabilistic outputs for structured questions.
  • Automating decision support in workflows like customer service, invoice processing, and security incident analysis.
  • Scenarios where understanding the confidence or likelihood of different options is crucial.
  • Developers seeking a compact model for typed decision tasks with competitive accuracy and low latency.

Finetunes
1 models