Fizzarolli/sappha-2b-v3

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:2.5BQuant:BF16Ctx Length:8kPublished:Mar 24, 2024License:gemmaArchitecture:Transformer0.0K Warm

Fizzarolli/sappha-2b-v3 is a 2.5 billion parameter instruction-tuned QLoRA fine-tune of the Gemma-2B base model, developed by Fizzarolli. This model, trained with Unsloth, demonstrates improved performance over its base model and Dolphin-2.8-Gemma-2B on MMLU, HellaSwag, and PIQA benchmarks. With an 8192-token context length, it is optimized for general instruction-following tasks.

Loading preview...

Overview

Fizzarolli/sappha-2b-v3 is an instruction-tuned model based on the Gemma-2B architecture, fine-tuned using QLoRA with Unsloth. This iteration, following a private v2 failure, aims to provide a more stable and capable small language model.

Key Capabilities & Performance

This model shows competitive performance against its base model, gemma-2b-it, and dolphin-2.8-gemma-2b across several benchmarks:

  • MMLU (five-shot): Achieves 38.02, outperforming both comparison models.
  • HellaSwag (zero-shot): Scores 51.70, also surpassing its counterparts.
  • PIQA (one-shot): Reaches 75.46, indicating strong common sense reasoning.
  • TruthfulQA (zero-shot): While slightly lower than gemma-2b-it, it remains competitive.

Prompt Format

The model utilizes a basic ChatML format for interactions, supporting clear system, user, and assistant turns.

Good For

  • General instruction-following tasks where a compact model size (2.5B parameters) is beneficial.
  • Applications requiring a model with an 8192-token context length.
  • Scenarios where improved benchmark performance over the base Gemma-2B model is desired.