metadata
license: gemma
library_name: transformers
pipeline_tag: image-text-to-text
extra_gated_heading: Access Gemma on Hugging Face
extra_gated_prompt: >-
To access Gemma on Hugging Face, you’re required to review and agree to
Google’s usage license. To do this, please ensure you’re logged in to Hugging
Face and click below. Requests are processed immediately.
extra_gated_button_content: Acknowledge license
base_model: google/gemma-3-27b-pt
Model Card for Dhruv-27b-preview
Model Overview
Dhruv-27B-preview is a powerful 27-billion parameter language model based on Gemma3, fine-tuned for advanced reasoning tasks. Dhruv-27B is part of the Gemini Reasoning Series, optimized for high performance on academic, logical, and factual evaluations. It is capable of deep contextual reasoning and chain-of-thought generation, making it suitable for research, enterprise applications, and AI agents requiring robust general knowledge understanding.
Key Metrics
Dhruv-27b-preview achieves strong performance on key benchmarks:
+------------------------+--------+
| Metric | Score |
+------------------------+--------+
| MMLU (5-shot) | 80.0% |
| GPQA (0-shot) | 52.0% |
+------------------------+--------+
Model Architecture
- Base Model: Gemma3
- Model Size: 27B parameters
- Type: Decoder-only Transformer (causal LM)
- Precision: bf16 with int8 quantization for inference efficiency
- Training Objective: Instruction-tuned with emphasis on reasoning, question answering, and factual correctness
Intended Use
- Research and academic QA tasks
- General-purpose reasoning agents
- Multilingual factual reasoning
- Enterprise AI tools requiring high factual accuracy and depth
USAGE
- Use a System Template
Limitations
- Not optimized for creative generation or dialog tasks.
- May hallucinate in areas with limited training data.
- Requires fine-grained prompt engineering for complex instructions.
Citation
@misc{vicharai_dhruv27b,
title={Dhruv-27B: Gemma3 Reasoning Model},
author={vicharai},
year={2025},
publisher={vicharai},
url={https://vichar.io}
}
Developed by vichar ai Huggingface Website