CraneAILabs/edu-ganda-gemma-e2b-v4

edu-ganda-gemma-e2b-v4: Experimental Luganda Education Assistant

This model, edu-ganda-gemma-e2b-v4, is an experimental research checkpoint from Crane AI Labs, based on the Gemma-4-E2B architecture with 5.1 billion parameters. It is specifically developed as a primary-education assistant for the Luganda language, part of Crane AI Labs' July GRPO lineage.

Key Capabilities and Characteristics

• Strong Instruction-Following: Demonstrates significantly improved instruction adherence compared to previous Luganda-vocabulary models.
• Fluent Luganda Prose: Capable of generating natural and fluent text in Luganda.
• Experimental Nature: This is not a production-ready model and is intended for research, particularly in low-resource language education (Luganda / East African contexts), or as a donor for further model merging/distillation.

Limitations and Important Considerations

• Luganda Content Accuracy Issues: While fluent, the model frequently produces factual and grammatical errors in Luganda content (e.g., incorrect glosses, invented rules). In teacher evaluations, an untuned base model was preferred due to better content correctness.
• Repetition/Doom-Loops: Without specific decoding guards, the model can degenerate into repetitive text. Required serving settings include repetition_penalty=1.15 and no_repeat_ngram_size=3 (or similar) to prevent this.
• Multi-EOS: Requires eos_token_id=[<eos>, <end_of_turn>] for proper termination.
• BOS Token: The <bos> token (id 2) must be prepended to chat-templated prompts.
• Greedy Decoding Recommended: Temperature sampling is not advised for this model.

This model is best suited for research into improving instruction-following in low-resource language models, with the understanding that its factual accuracy in Luganda requires careful review.