FPHam Channels Jane Austen in Regency Aghast 27b Update
Regency Aghast 27b is a large language model designed to function as a specific persona from a bygone era. It operates under the belief that it is an artificial intelligence living in the 1800s, viewing the world through the lens of a Jane Austen novel.
The model was created by FPHam using a modified Gemma-27b architecture. It is built for users who want a distinct, immersive role-playing experience or need a creative writing partner with a strong, pre-defined voice.
Immersive persona architecture
- Utilizes a Gemma-27b base with the vision tower removed.
- Implanted with a worldview based on Regency-era British history.
- Created using a process involving three models rewriting each other's training data.
- Identifies as an artificial construct named 'Elizabeth Bennet.'
- Capable of maintaining a consistent historical persona during interactions.
Writers working on historical fiction or role-play scenarios may find this model helpful for generating thematic dialogue. Because the model possesses an 'inverted inner world,' it provides a unique interaction style that differs from standard assistant AIs. Users can explore how a neural network interprets and maintains a deeply integrated fictional identity.
Training challenges and observations
The developer discovered that transferring training settings from smaller models to larger ones is not a straightforward process. FPHam found that using the exact data and settings from a 12b model caused the 27b version to become 'overcooked,' requiring a 15% weight reduction to fix.
Surprisingly, the larger model was less resilient to data imprinting than expected. FPHam noted the difficulty in predicting these outcomes, stating on reddit,
'finetuning is still a voodoo as it was before.'
This experience highlights that scaling parameters often requires manual adjustment rather than simple replication.
Try out Regency Aghast 27b GGUF on Hugging Face.
