Merlin Model Card

Model Details

Model Name: Merlin

Model Type: Transformer-based leveraging Microsoft Phi 14b 128k tokens

Publisher: Awels Engineering

License: MIT

Model Description: Merlin is a sophisticated model designed to help as an AI agent focusing on the Druid AI Conversational platform. It leverages advanced machine learning techniques to provide efficient and accurate solutions. It has been trained on the full docments corpus of Druid 7.14.

Dataset

Dataset Name: awels/druidai_admin_dataset

Dataset Source: Hugging Face Datasets

Dataset License: MIT

Dataset Description: The dataset used to train Merlin consists of all the public documents available on the Druid AI Conversational Platform. This dataset is curated to ensure a comprehensive representation of typical administrative and development scenarios encountered in Druid AI Platform.

Training Details

Training Data: The training data includes 33,000 Questions and Answers generated by the Bonito LLM. The dataset is split into 3 sets of data (training, test and validation) to ensure robust model performance.

Training Procedure: Thready was trained using supervised learning with cross-entropy loss and the Adam optimizer. The training involved 1 epoch, a batch size of 4, a learning rate of 5.0e-06, and a cosine learning rate scheduler with gradient checkpointing for memory efficiency.

Hardware: The model was trained on a single NVIDIA H100 SXM graphic card.

Framework: The training was conducted using PyTorch.

Evaluation

Evaluation Metrics: Thready was evaluated on the training dataset:

epoch = 1.0 total_flos = 124998759GF train_loss = 1.8515 train_runtime = 0:43:52.83 train_samples_per_second = 9.584 train_steps_per_second = 2.396

Performance: The model achieved the following results on the evaluation dataset:

epoch = 1.0 eval_loss = 1.5167 eval_runtime = 0:01:56.08 eval_samples = 5298 eval_samples_per_second = 52.287 eval_steps_per_second = 13.076

Intended Use

Primary Use Case: Merlin is intended to be used locally in an agent swarm to colleborate together to solve Druid AI Conversational platform related problems.

Limitations: This 14b model is an upscale of the 3b model. Much better loss than the 3b so results should be better.