Kuwutu-7B-CYOA
This is the merged full-model version of
KeinNiemand/Kuwutu-7B-CYOA-LoRA, a narrow fine-tune of
Mawdistical/Kuwutu-7B for adult furry CYOA / interactive-fiction style continuation.
It is not a general writing model. It was trained mostly on a narrow permissioned CYOA-style source and should be treated as a domain/style fine-tune.
Other Versions
Intended Use
- Fictional adult furry CYOA / interactive-fiction continuation.
- Long-form scene continuation from ChatML-style prompts.
- Branching narrative drafts and choice-driven writing experiments.
Not Intended For
- General assistant use.
- Factual, medical, legal, or safety-critical advice.
- Content involving minors.
- Real-person sexual content without consent.
- Harassment, impersonation, or non-consensual sexual scenarios.
Training Notes
- Base model: Mawdistical/Kuwutu-7B
- Adapter source: KeinNiemand/Kuwutu-7B-CYOA-LoRA
- Training method: LoRA SFT, merged into the base model
- Checkpoint:
run0067, checkpoint 274
- Rank:
32
- Alpha:
32
- LoRA dropout:
0
- Weight decay:
10
- Optimizer: Muon with AdamW fallback parameters
- Template: ChatML
The checkpoint was selected as a pragmatic loss/guardrail compromise. Generation judge results were noisy in testing and were not used as the sole release signal.
Data
The training data is permissioned adult furry CYOA / interactive-fiction writing. The dataset is not redistributed here.
Limitations
This is an experimental narrow-domain model. It may produce repetitive, generic, incoherent, overly specific, or poorly structured continuations. It may fail at scene continuity, choice structure, lore consistency, or stopping behavior.
