Dhenu2 India 8B

Model Overview

Model Name: Llama3.1-Dhenu2-In-8B-Instruct

Architecture: Llama3.1

Parameters: 8 Billion

Release Date: 24th October, 2024

License: Llama 3.2 Community License

Description

Dhenu2 India 8B is our flagship agricultural language model, meticulously trained on the Llama3.1 architecture. Optimized for India's diverse agricultural practices, it delivers actionable insights and knowledgeable responses tailored to the unique needs of Indian farmers, policymakers, and agri-businesses. This model is ideal for developing comprehensive advisory applications that support informed decision-making in the agricultural sector.

Intended Use

• Advisory Applications: Build tools that provide farmers with real-time advice on crop management, pest control, package of practices, and resource optimization.

Training Data

Dhenu2 India 8B was trained on a diverse dataset comprising:

• Instruction Set: Over 1.5 million instructions from real and synthetic conversations.
• Synthetic Instructions: Generated through advanced pipelines to cover more than 4,000 agricultural topics.
• Data Sources: Mobile extension service logs, farmer feedback, agricultural package of practices, and localized studies.

Training Procedure

• Techniques: Full fine-tuning combined Adaptation (LoRA with Low-Rank) to optimize performance while managing computational resources.
• Hardware: Trained using multi-GPU setups with NVIDIA A100 GPUs, leveraging DeepSpeed for distributed training and memory optimization.
• Optimization: Implemented flash attention mechanisms to enhance computational efficiency and reduce memory overhead.

Evaluation

• Human Evaluation: Assessed by agricultural experts for relevance, accuracy, and actionable insights.
• Synthetic Evaluation: Peer-reviewed by other Large Language Models (LLMs) to ensure consistency and correctness.
• Performance Metrics: Evaluated based on precision, recall, and domain-specific accuracy in delivering agricultural insights.

Limitations

While Dhenu2 India 8B excels in agricultural contexts, it may not perform as effectively outside this domain. Users should ensure that applications leveraging this model are contextually relevant to agriculture to maintain response accuracy and reliability.

API

Use our platform Dhenu with a generous free quota to start building your agriculture applications.

A note of gratitude

We want to thank our partners Microsoft and Microsoft for Startups for landing us compute. We would also like to thank our partner, Meta, for the open-source Llama models.

Contact

For more information, support, or collaboration inquiries, please contact us at [support@kissan.ai].