Overview

This model, activeDap/gemma-2b_ultrafeedback_chosen, is a fine-tuned variant of Google's Gemma-2b base model, featuring 2.5 billion parameters. It has been specifically trained using Supervised Fine-Tuning (SFT) on the activeDap/ultrafeedback_chosen dataset. The training process involved 826 steps, achieving a final training loss of 1.6215 over 1 epoch.

Key Capabilities

• Instruction Following: Optimized for generating coherent and relevant responses based on given prompts, leveraging the ultrafeedback_chosen dataset's structure.
• Assistant-Style Generation: Trained with an Assistant-only loss, making it proficient in producing helpful and conversational outputs.
• Efficient Inference: As a 2.5 billion parameter model, it offers a balance between performance and computational efficiency.

Training Details

The model was trained with a per-device batch size of 16, accumulating gradients over 4 GPUs for a total batch size of 64. It utilized a cosine learning rate scheduler with a 0.1 warmup ratio and a maximum sequence length of 512 tokens. The training framework involved Transformers and TRL libraries.

Good For

• Developing conversational agents and chatbots.
• Instruction-tuned text generation tasks.
• Applications requiring a compact yet capable language model for assistant-like interactions.