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.
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|>
