uhhlt/story-emb
uhhlt/story-emb is a 7 billion parameter embedding model, based on the intfloat/e5-mistral-7b-instruct architecture, specifically fine-tuned for generating narrative-focused representations of fictional stories. This model excels at story retrieval tasks, allowing users to find stories with similar narrative structures. It was developed by Hatzel and Biemann, and is optimized for understanding and comparing narrative content.
Loading preview...
StoryEmb: Narrative-Focused Story Embeddings
StoryEmb is a 7 billion parameter embedding model, derived from the intfloat/e5-mistral-7b-instruct architecture, and specifically trained to create narrative-focused representations of fictional stories. This model is the result of research by Hatzel and Biemann, detailed in their 2024 paper "Story Embeddings — Narrative-Focused Representations of Fictional Stories" presented at ACL.
Key Capabilities
- Narrative-Focused Embeddings: Generates embeddings that capture the narrative structure and content of fictional stories.
- Augmented Data Training: The published variant is trained on augmented data for enhanced performance.
- Compatibility: Can be used similarly to the base
intfloat/e5-mistral-7b-instructmodel.
Primary Use Case
- Story Retrieval: Optimized for retrieving stories with similar narrative structures. Users should apply the task prefix "Retrieve stories with a similar narrative to the given story: " for optimal results.
Additional Information
- Adapter weights are available in the
adapter-weightsdirectory for those interested in continued fine-tuning. - The model's development is thoroughly documented in the associated research paper.