Athena-Class Model B: Dataset and Reproducibility Artifacts

This record provides the specialized fine-tuning dataset for Athena-Class Model B and related artifacts for reproducibility.

Included in this record:

• Athena-Class Model B multimodal training dataset
• Fine-tuning script

If you intend to reproduce the Model B, it is critical that you understand the parameters included in the training script.

A few key points to understand:

• This is not a typical configuration, particularly for a Rank 512 / Alpha 1024 fine-tune
• 16K max sequence length is necessary to accomodate each training segment (system prompt + data + model output)
• Use the included modified jinja that supports multiple images in each training segment and addresses the known thinking template isues placing {%- set enable_thinking = true -%} at the top
• The model's output is included in the fine-tune (no SFT)