Writer-Large-2411-v2.1

EXL2-Quant available here: gghfez/Writer-Large-2411-v2.1-exl2-4.5bpw

Creative-Writing Control-Vectors available here: gghfez/Writer-Large-2411-v2.1-control-vectors

Overview

This model is built on Mistral-Large-Instruct-2411 and optimized for creative writing purposes. The base model excels at following instructions and handling details in long context when using the new prompt template.

Key Improvements

• Reduced positivity bias
• Reduced AI tropes and repetitive language patterns in story generation
• Enhanced performance with longer context stories (multiple chapters) and roleplay sessions
• Improved steering capabilities for roleplay via [OOC] instructions
• Better handling of "group chat" scenarios

Usage

Prompt Template

The model works best with a system prompt in the Mistral-V7 format. If you omit [SYSTEM_PROMPT] [/SYSTEM_PROMPT], the model:

• May not follow instructions properly at short contexts
• Can become repetitive at longer contexts

Example:

[SYSTEM_PROMPT]You are an award winning writer. Assist the user.[/SYSTEM_PROMPT][INST] Write the opening chapter of ... [/INST]

SillyTavern Integration

With System Prompt:

Story String:

[SYSTEM_PROMPT] {{#if system}}{{system}}[/SYSTEM_PROMPT] [INST]
{{/if}}{{#if wiBefore}}{{wiBefore}}
{{/if}}{{#if description}}{{description}}
{{/if}}{{#if personality}}{{personality}}
{{/if}}{{#if scenario}}{{scenario}}
{{/if}}{{#if wiAfter}}{{wiAfter}}
{{/if}}{{#if persona}}{{persona}}
{{/if}}{{trim}}[/INST] Understood.</s>

Without System Prompt:

Story String:

[INST]{{#if system}}{{system}}[/SYSTEM_PROMPT]
{{/if}}{{#if wiBefore}}{{wiBefore}}
{{/if}}{{#if description}}{{description}}
{{/if}}{{#if personality}}{{personality}}
{{/if}}{{#if scenario}}{{scenario}}
{{/if}}{{#if wiAfter}}{{wiAfter}}
{{/if}}{{#if persona}}{{persona}}
{{/if}}{{trim}}[/INST] Understood.</s>

For response steering, use [OOC] commands, e.g.:

• [OOC] Have them interrupted by a loud explosion in a nearby factory
• [OOC] Have her refuse to sell it and suggest another merchant instead

Technical Details

Training

• QLoRA training at 32768 context
• Merged with gghfez/Mistral-Large-Instruct-2411 at bf16
• jukofyork/Creative writing control vectors were applied during synthetic dataset generation
• Includes standard assistant instruct data for long-context stability
• Note: Performance on code tasks may be reduced compared to base model
• Note: No attempt was made to remove 'Name-Slop', so you'll still encounter Lily and Elara if you don't specify character names

Context Length

• Base model: 131,072 tokens
• Training range: 1024-32728 tokens
• Training context window: 32768 tokens

Testing Environments

Tested with exllamav2 4.5bpw on:

• tabbyAPI + MikuPad
• tabbyAPI + SillyTavern
• exui