jspr/talosian-7b

Talosian-7B: Controllable Story Generation

Talosian-7B is a 7 billion parameter language model developed by jspr, built upon the Mistral-7B v0.2 base architecture. Its core purpose is controllable storytelling, allowing users to generate narratives section-by-section. The model was fine-tuned on a long-context dataset of fanfiction stories, which underpins its ability to produce creative and coherent narrative segments.

Key Capabilities

• Section-by-Section Story Generation: Talosian-7B is uniquely designed to write stories in discrete sections, guided by specific "section headers" (e.g., [SEC]...[/SEC]). This enables granular control over the narrative flow.
• Structured Prompting: It utilizes a distinct prompting format where section headers are interspersed with story content, allowing for detailed guidance on the action within each segment. An [END] token within the final section header helps the model conclude stories naturally.
• Creative Writing Focus: Optimized for generating imaginative and engaging story content, particularly suited for fanfiction or similar narrative forms.
• Mistral-7B v0.2 Foundation: Inherits the robust capabilities of the Mistral-7B v0.2 base model, including a 32k token context length.

Good For

• Interactive Storytelling: Developers building applications where users guide story progression through prompts.
• Creative Writing Assistance: Aiding writers in structuring and generating specific parts of a story.
• Fanfiction Generation: Producing new fanfiction content with controlled plot points.

Talosian-7B is intended for completion-style prompting (e.g., in text-generation-webui's Notebook mode) rather than traditional chat formats. Optimal generation parameters include a temperature between 0.5 and 0.9, and [SEC] can be used as a custom stopping string for single-section generation.