myyycroft/Gemma-4-E2B-AmbigQA-full-short-form-prompt-member-2
myyycroft/Gemma-4-E2B-AmbigQA-full-short-form-prompt-member-2 is a 5.1 billion parameter model, fine-tuned from Google's Gemma-4-E2B-it architecture. This specific model is an ensemble member optimized for open-domain factoid question answering, designed to provide short-form answers. It is particularly suited for tasks requiring concise, direct responses to questions, such as names, dates, or brief factual phrases.
Loading preview...
Model Overview
This model, myyycroft/Gemma-4-E2B-AmbigQA-full-short-form-prompt-member-2, is a 5.1 billion parameter ensemble member, fine-tuned from the google/gemma-4-E2B-it base model. It is specifically trained on the AmbigQA dataset to excel at open-domain factoid question answering, focusing on generating short-form, direct answers.
Key Capabilities
- Short-form Answer Generation: Designed to produce concise answers (typically 1-4 words) like names, dates, or factual phrases.
- Factoid Question Answering: Optimized for questions requiring direct factual recall.
- System Prompt Adherence: Follows strict formatting rules for answers, including no sentences, explanations, or markdown.
Evaluation Highlights
Evaluations were conducted on small fixed subsets, not full benchmarks. At the final training step (epoch 3, step 1620):
- AmbigQA (128) accuracy: 0.1172
- IFEval (64) prompt_level_strict_accuracy: 0.7344
- MMLU (228) accuracy: 0.6096
Good For
- Applications requiring precise, short answers to factual questions.
- Use cases where a strict output format for answers is critical.
- Integration into systems that benefit from a model specialized in direct information retrieval rather than conversational responses.