CraneAILabs/edu-ganda-gemma-e2b-v4

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:5.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 3, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

CraneAILabs/edu-ganda-gemma-e2b-v4 is a 5.1 billion parameter experimental research checkpoint from Crane AI Labs, based on the Gemma-4-E2B architecture. This model is specifically designed as a Luganda primary-education assistant, optimized through continued pretraining on distilled Luganda dictionary documents, translation training, and pedagogy fine-tuning. It excels in Luganda language generation and numeracy tasks, making it suitable for educational applications requiring Luganda language support.

Loading preview...

Overview

CraneAILabs/edu-ganda-gemma-e2b-v4 is an experimental research checkpoint developed by Crane AI Labs, serving as a parent model for the final EduGanda-Gemma-4-E2B-v5. This 5.1 billion parameter model is built upon the Gemma-4-E2B architecture and is specifically engineered to function as a primary-education assistant for the Luganda language.

Key Capabilities & Training

This model underwent a specialized training regimen including:

  • Continued pretraining on distilled Luganda dictionary documents.
  • Translation training for Luganda to English and English to Luganda, demonstrating significant improvements over the base Gemma-4-E2B model (e.g., chrF++ scores of 48.9 / 44.1 compared to 17.6 / 14.6).
  • Curriculum and pedagogy fine-tuning with constraint-following examples.
  • Two reinforcement-learning rounds to enhance performance.
  • Quantisation-aware training (QAT) for efficient serving.

Performance Highlights

Compared to the base Gemma-4-E2B model, edu-ganda-gemma-e2b-v4 shows notable improvements in:

  • Numeracy: Achieves 73.3% on held-out problems, significantly higher than the base model's 48.7%.
  • Luganda Quality: Judged Luganda quality is 3.90 out of 10, a substantial increase from the base model's 0.28, particularly in tasks requiring Luganda output.
  • Instruction Adherence: Shows 63.6% adherence on 220 program-checked prompts.

Limitations

  • Luganda Content Accuracy: May produce factual or grammatical errors in Luganda content; human review is recommended.
  • Repetitive Output: Can loop on drill-style prompts without a repetition guard; repetition_penalty=1.0 with no_repeat_ngram_size=4 is recommended.

Recommended Use Cases

This model is particularly well-suited for:

  • Developing Luganda primary-education tools and assistants.
  • Translation tasks involving Luganda and English.
  • Applications requiring numeracy support within a Luganda context.

Finetunes
1 models