myyycroft/Gemma-4-E2B-AmbigQA-full-member-2
The myyycroft/Gemma-4-E2B-AmbigQA-full-member-2 is a 5.1 billion parameter language model, fine-tuned from Google's Gemma-4-E2B-it architecture. This specific model is an ensemble member (seed 2060) optimized for question answering on the AmbigQA dataset. It demonstrates capabilities in handling ambiguous questions and general language understanding, with a context length of 32768 tokens.
Loading preview...
Model Overview
This model, myyycroft/Gemma-4-E2B-AmbigQA-full-member-2, is a 5.1 billion parameter language model derived from the google/gemma-4-E2B-it architecture. It represents the second member (seed 2060) of a five-member ensemble, fine-tuned specifically on the AmbigQA dataset to enhance its performance in answering ambiguous questions.
Key Capabilities & Training
- AmbigQA Specialization: Fine-tuned on the
sewon/ambig_qadataset, this model is designed to address questions that may have multiple valid answers or require clarification. - Ensemble Member: It is part of a larger ensemble, contributing to a more robust overall system, with its training configured for 3 epochs and 1620 steps.
- Evaluation Metrics: On small, fixed subsets, this member achieved an AmbigQA accuracy of 0.0938 and an AlignScore of 0.1924. It also showed an IFEval strict accuracy of 0.6719 and MMLU accuracy of 0.5702.
- Hyperparameters: Utilizes a learning rate of 2.0e-05, a batch size of 4 per device, and a maximum sequence length of 512, with gradient checkpointing enabled.
Use Cases
This model is particularly suited for applications requiring nuanced question answering, especially in scenarios where queries might be ambiguous or open to multiple interpretations. Its fine-tuning on AmbigQA suggests its strength in handling complex information retrieval and response generation.