96 lines
3.6 KiB
Markdown
96 lines
3.6 KiB
Markdown
---
|
|
base_model: Mawdistical/Kuwutu-7B
|
|
base_model_relation: finetune
|
|
license: cc-by-nc-4.0
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
language:
|
|
- en
|
|
tags:
|
|
- qwen2
|
|
- chatml
|
|
- interactive-fiction
|
|
- cyoa
|
|
- roleplay
|
|
- creative-writing
|
|
- furry
|
|
- anthro
|
|
- nsfw
|
|
- explicit
|
|
- not-for-all-audiences
|
|
---
|
|
|
|
# Kuwutu-7B-CYOA-v2
|
|
|
|
### Explicit Content Warning
|
|
|
|
Kuwutu-7B-CYOA-v2 is a fine-tune of [Mawdistical/Kuwutu-7B](https://huggingface.co/Mawdistical/Kuwutu-7B?not-for-all-audiences=true) for CYOA-style interactive fiction in the adult furry, dark, and roleplay-oriented themes of the base model.
|
|
|
|
---
|
|
|
|
## ✧ Versions
|
|
|
|
- **LoRA adapter:** [Kuwutu-7B-CYOA-LoRA-v2](https://huggingface.co/KeinNiemand/Kuwutu-7B-CYOA-LoRA-v2?not-for-all-audiences=true)
|
|
- **Adapter GGUF:** [Kuwutu-7B-CYOA-LoRA-GGUF-v2](https://huggingface.co/KeinNiemand/Kuwutu-7B-CYOA-LoRA-GGUF-v2?not-for-all-audiences=true)
|
|
- **Mainline GGUF:** [Kuwutu-7B-CYOA-GGUF-v2](https://huggingface.co/KeinNiemand/Kuwutu-7B-CYOA-GGUF-v2?not-for-all-audiences=true)
|
|
- **IK GGUF:** [Kuwutu-7B-CYOA-IK-GGUF-v2](https://huggingface.co/KeinNiemand/Kuwutu-7B-CYOA-IK-GGUF-v2?not-for-all-audiences=true)
|
|
|
|
---
|
|
|
|
## ✧ What Changed in v2
|
|
|
|
v2 is a retrained release with substantially better results than v1 on the private generation benchmark and automated loop detection.
|
|
|
|
## ✧ Intended Use
|
|
|
|
This model is intended for long-form interactive fiction, especially CYOA-style scene continuation and branching narrative in the themes of Kuwutu-7B.
|
|
|
|
It uses the ChatML template. Prompts can provide the current scene, player state, relevant context, and instructions for the next continuation.
|
|
|
|
## ✧ Training Details
|
|
|
|
| Setting | Value |
|
|
|---|---|
|
|
| Base model | `Mawdistical/Kuwutu-7B` |
|
|
| Method | LoRA SFT, merged for this repository |
|
|
| Release checkpoint | `run0072`, checkpoint `282` |
|
|
| Rank / alpha | `16 / 16` |
|
|
| LoRA+ ratio | `4.6` |
|
|
| Optimizer | Muon with AdamW fallback parameters |
|
|
| Learning rate | `2.2e-4` |
|
|
| Weight decay | `10` |
|
|
| Epochs | `2` |
|
|
| Context cutoff | `15,872` tokens |
|
|
| Precision | BF16 |
|
|
| Chat template | ChatML |
|
|
|
|
The training data consists of permissioned adult CYOA / interactive-fiction writing and a small amount of related long-form writing. The dataset is not redistributed with the model.
|
|
|
|
## ✧ Evaluation
|
|
|
|
Results from a private pairwise LLM-judged benchmark of generated continuations:
|
|
|
|
| Candidate | Elo-like score | Catastrophic-loop rate |
|
|
| :--- | ---: | ---: |
|
|
| Real | 510.9 ± 46.9 | 0/114 (0.0%) |
|
|
| Kuwutu-7B-CYOA-v2 | 72.7 ± 37.3 | 0/112 (0.0%) |
|
|
| Kuwutu-7B | 4.0 ± 31.8 | 0/114 (0.0%) |
|
|
| Kuwutu-7B-CYOA-v1 | -46.6 ± 34.1 | 16/112 (14.3%) |
|
|
| Dummy | -423.1 ± 45.6 | 0/114 (0.0%) |
|
|
|
|
These private internal results should be treated as directional rather than as broad capability benchmarks.
|
|
|
|
## ✧ Limitations
|
|
|
|
The model may hallucinate story facts, drift from scene context, repeat phrases, overproduce prose, or mishandle choice structure. Its behavior remains strongly influenced by Kuwutu-7B and the narrow interactive-fiction training distribution.
|
|
|
|
## ✧ License
|
|
|
|
Released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/), following the base model license. Attribution is required and commercial use is not permitted under this license.
|
|
|
|
## ✧ Credits
|
|
|
|
- **Base model:** [Mawdistical/Kuwutu-7B](https://huggingface.co/Mawdistical/Kuwutu-7B?not-for-all-audiences=true)
|
|
- **Training / export:** [LLaMA Factory](https://github.com/hiyouga/LLaMA-Factory)
|
|
- **GGUF tooling:** [llama.cpp](https://github.com/ggml-org/llama.cpp) and [ik_llama.cpp](https://github.com/ikawrakow/ik_llama.cpp)
|