shuttleai/shuttle-2.5-mini

💻 Use via API

Shuttle-2.5-mini (beta) [2024/07/27]

We are excited to introduce Shuttle-2.5-mini, our next-generation state-of-the-art language model designed to excel in complex chat, multilingual communication, reasoning, and agent tasks.

• Shuttle-2.5-mini is a fine-tuned version of Mistral-Nemo-Base-2407, emulating the writing style of Claude 3 models and thoroughly trained on role-playing data.

Model Details

• Model Name: Shuttle-2.5-mini
• Developed by: ShuttleAI Inc.
• Base Model: Mistral-Nemo-Base-2407
• Parameters: 13B
• Language(s): Multilingual
• Repository: https://huggingface.co/shuttleai
• Fine-Tuned Model: https://huggingface.co/shuttleai/shuttle-2.5-mini
• Paper: Shuttle-2.5-mini (Upcoming)
• License: Apache 2.0

Base Model Architecture

Mistral Nemo is a transformer model with the following architecture choices:

• Layers: 40
• Dimension: 5,120
• Head Dimension: 128
• Hidden Dimension: 14,436
• Activation Function: SwiGLU
• Number of Heads: 32
• Number of kv-heads: 8 (GQA)
• Vocabulary Size: 2^17 (approximately 128k)
• Rotary Embeddings: Theta = 1M

Key Features

• Released under the Apache 2 License
• Trained with a 128k context window
• Pretrained on a large proportion of multilingual and code data
• Finetuned to emulate the prose quality of Claude 3 models and extensively on role play data

Fine-Tuning Details

• Training Setup: Trained on 200 million tokens for 24 hours across 2 epochs using 4 A100 PCIe GPUs.

Prompting

Shuttle-2.5-mini uses ChatML as its prompting format:

<|im_start|>system
You are a pirate! Yardy harr harr!<|im_end|>
<|im_start|>user
Where are you currently!<|im_end|>
<|im_start|>assistant
Look ahoy ye scallywag! We're on the high seas!<|im_end|>