locailabs/gemma-3-1b-it-sft-metamathqa-modelmerge
The locailabs/gemma-3-1b-it-sft-metamathqa-modelmerge is a 3.1 billion parameter Gemma-based instruction-tuned language model developed by locailabs. This model is created by merging a MetaMathQA-finetuned version with the original base model using linear interpolation. It is specifically optimized for mathematical reasoning and instruction following, demonstrating improved performance on tasks like GSM8K compared to its base and purely fine-tuned counterparts.
Loading preview...
Model Overview
This model, developed by locailabs, is a Gemma 3.1B instruction-tuned variant specifically enhanced for mathematical reasoning. It was created using a unique model merging technique where a version fine-tuned on the MetaMathQA dataset was linearly interpolated with the original base model (α=0.5).
Key Capabilities & Performance
The merging approach aims to improve task-specific gains from fine-tuning while simultaneously retaining the base model's general instruction-following abilities, which can sometimes degrade with pure fine-tuning. Benchmarking results highlight this strategy:
- GSM8K: The merged model achieves 39.58, outperforming both the base (33.66) and the purely fine-tuned (37.15) versions.
- MMLU Redux: Scores 40.53, showing a balanced performance between the base (39.79) and FT (41.02).
- IFEval: Demonstrates a score of 36.41, significantly better than the fine-tuned model's 28.84, indicating better instruction adherence.
Good For
- Applications requiring enhanced mathematical problem-solving capabilities.
- Scenarios where strong instruction following is crucial alongside specialized task performance.
- Developers looking for a compact model with balanced performance across reasoning and general instruction tasks.