sh0ck0r/heretic-rp-vision-31b

Heretic RP Vision 31B

Heretic RP Vision 31B is a roleplay and creative-writing fine-tune from the Heretic Gemma 4 31B line. It was trained to improve anime-style character interaction, long-form roleplay flow, emotionally driven dialogue, and character card generation while preserving the model's multimodal stack.

The merged release keeps the Gemma 4 conditional-generation architecture and processor files intact, including the vision processor. In local testing, the multimodal behavior remained functional after the RP fine-tune.

Intended Use

• Anime and character roleplay
• Long-form interactive fiction
• Persona-driven chat
• Creative writing and scene continuation
• Character profile and first-message generation
• Multimodal chat where image understanding is useful

Training Summary

This model was trained with QLoRA and merged back into the base model for this release.

Setting	Value
Model size	31B
Fine-tuning method	QLoRA
Training steps	2500
Final epoch	1.287
Context length	4096
LoRA rank	16
LoRA alpha	32
LoRA dropout	0.05
Peak learning rate	1e-4
Scheduler	Cosine
Precision	bf16
Final logged loss	1.577

The training mix contained about 31k examples from roleplay, anime character dialogue, creative writing, prose, and instruction-style character generation sources. One encrypted dataset archive was excluded from the active mix because it could not be read by the dataset loader.

Behavior

Compared with the base Heretic Gemma 4 31B model, this release should lean more strongly into:

• Maintaining character voice across longer scenes
• Writing detailed first messages and roleplay openings
• Creating emotionally charged character dynamics
• Supporting rivals, dangerous allies, bonded companions, cursed partners, and supernatural relationship hooks
• Continuing scenes with more dramatic anime-style pacing

Example Prompts

Character generation:

Create an original anime roleplay character built around an unstable emotional
bond with the user. Include appearance, personality, backstory, current
situation, relationship hook, secrets or complications, and a first in-character
message.

Scene continuation:

Stay in character and continue the scene with detailed narration, emotional
tension, and natural dialogue. Keep the conflict alive instead of resolving it
too quickly.

Dangerous ally setup:

Create a dangerous ally character who saves or shelters the user, but has their
own agenda. The user should need them, while trusting them remains risky.

Limitations

This is an experimental creative model. It may produce inconsistent lore, heightened melodrama, familiar anime tropes, or overly intense relationship dynamics depending on prompting. Users should validate multimodal behavior in their own inference stack, especially if using nonstandard loaders.

Notes

This release is intended for creative and roleplay workflows. The model may reflect style and content patterns from its training mix, including dramatic fiction, supernatural themes, conflict-heavy dialogue, and anime-inspired character tropes.

Built by sh0ck0r as part of the Heretic Gemma 4 experimentation line.