File size: 7,745 Bytes

---
license: apache-2.0
language:
  - en
tags:
  - renpy
  - visual-novel
  - storytelling
  - creative-writing
  - llama-3
  - qlora
  - finetuned
  - text-generation
  - natural-language-to-script
  - instruction-free
inference: false
widget:
  - text: "A detective wakes up in a town where no one remembers him."
  - text: "Every time she falls asleep, she wakes up in another version of her life."
  - text: "A cursed mirror swaps your life with your reflection."
datasets:
  - custom
model-index:
  - name: Secunda-0.3-F16-QA
    results: []
---


```ascii
                          
       ▄▄▄▄▄   ▄███▄   ▄█▄      ▄      ▄   ██▄   ██       ████▄     ▄█ 
       █     ▀▄ █▀   ▀  █▀ ▀▄     █      █  █  █  █ █      █   █     ██ 
    ▄   ▀▀▀▀▄   ██▄▄    █   ▀  █   █ ██   █ █   █ █▄▄█     █   █     ██ 
     ▀▀▄▄▄▄▀    █▄   ▄▀ █▄  ▄▀ █   █ █ █  █ █  █  █  █     █   █     ▐█ 
               ▀███▀   ▀███▀  █▄ ▄█ █  █ █ ███▀     █     ▀████  ▐█  ▐   
                               ▀▀▀  █   ██         █                          
                                                 ▀                           
   
                      ⋆⋆୨୧˚ THE PRIMÉTOILE ENGINE ˚୨୧⋆｡˚⋆
                  — Visual Novel generation under starlight —
```
| Version                                               | Type            | Strengths                                                                 | Weaknesses                                                           | Recommended Use             |
|-------------------------------------------------------|-----------------|---------------------------------------------------------------------------|----------------------------------------------------------------------|-----------------------------|
| [Secunda-0.1-GGUF](https://huggingface.co/Yaroster/Secunda-0.1-GGUF) / [RAW](https://huggingface.co/Yaroster/Secunda-0.1-RAW) | Instruction     | - Most precise<br>- Coherent code<br>- Perfected Modelfile                | - Smaller context / limited flexibility                              | **Production / Baseline**  |
| [Secunda-0.3-F16-QA](https://huggingface.co/Yaroster/Secunda-0.3-F16-QA)        | QA-based Input  | - Acceptable for question-based generation                               | - Less accurate than 0.1<br>- Not as coherent                        | Prototyping (QA mode)       |
| [Secunda-0.3-F16-TEXT](https://huggingface.co/Yaroster/Secunda-0.3-F16-TEXT)    | Text-to-text    | - Flexible for freeform tasks                                            | - Slightly off<br>- Modelfile-dependent                             | Experimental / Text rewrite |
| [Secunda-0.3-GGUF](https://huggingface.co/Yaroster/Secunda-0.3-GGUF)            | GGUF build      | - Portable GGUF of 0.3                                                   | - Inherits 0.3 weaknesses                                            | Lightweight local testing   |
| [Secunda-0.5-RAW](https://huggingface.co/Yaroster/Secunda-0.5-RAW)              | QA Natural      | - Best QA understanding<br>- Long-form generation potential               | - Inconsistent output length<br>- Some instability                   | Research / Testing LoRA     |
| [Secunda-0.5-GGUF](https://huggingface.co/Yaroster/Secunda-0.5-GGUF)            | GGUF build      | - Portable, inference-ready version of 0.5                                | - Shares issues of 0.5                                               | Offline experimentation     |
| [Secunda-0.1-RAW](https://huggingface.co/Yaroster/Secunda-0.1-RAW)              | Instruction     | - Same base as 0.1-GGUF                                                   | - Same as 0.1                                                        | Production backup           |

---

## 🌙 Overview

**Secunda-0.1-RAW** is the original release of the Secunda fine-tuned model family, trained to produce polished **Ren'Py `.rpy` scripts** from structured instructions!

The model outputs:

* `define` blocks for named characters (with colors!)
* `image` declarations for scenes & sprites
* A clear `label start:` structure
* Emotional dialogue, branching `menu`s, `jump`s, and proper `return`

This version is *the most stable so far* — often more reliable than 0.3!

---

<b>/!\ NO HUMAN-MADE DATA WAS USED TO TRAIN THIS AI !
Secunda takes much pride in making sure the training data is scripted ! /!\ </b>

If you like Visual Novels, please visit [itch.io](itch.io) and support independant creators ! 


## ✨ Moonlight Specs

* **Base model**: `meta-llama/Meta-Llama-3.1-8B`
* **Fine-tuning**: QLoRA (r=64, alpha=16, dropout=0.1)
* **Precision**: Float16 (FP16)
* **Max tokens**: 4096
* **Hardware used**: RTX 4070, 64GB RAM

---

## 🪄 Inference in the Starlight

### Setup

## 🚀 Quick Start

### Installation

```bash
pip install transformers accelerate peft bitsandbytes
```

### Inference Script Example

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

BASE_MODEL = "meta-llama/Meta-Llama-3.1-8B"
LORA_PATH = "path/to/Secunda-0.1-RAW"

model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(model, LORA_PATH)
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

def build_prompt(idea):
    return f"""You are an expert writer of visual novels in Ren'Py. 
Generate a complete and polished Ren'Py script based on the following concept:

\"\"\"{idea}\"\"\"

Your output should include:
- `define` blocks for all characters (with names and color codes)
- `image` blocks for key backgrounds and character sprites
- `label start:` with a clear beginning
- Proper `scene`, `show`, `menu`, `play music/sound`, and `jump` statements
- Emotional dialogue and natural pacing
- A proper ending (`return`) or narrative closure

Structure the script as a `.rpy` file — do not include explanations, comments, or placeholder text."""

prompt = build_prompt("A young girl finds a photo album that shows moments that haven't happened yet.")
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.85, top_p=0.95)
    print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```


---

## 🌌 Evaluation

This model has:

* Generated 1000+ `.rpy` files
* Passed human review for structure, creativity & syntax
* > 90% valid output with minimal manual tweaks

---

## ☁️ Talking to the Moon

If you use **Secunda-0.1-RAW**, please star and cite:

```bibtex
@misc{secunda2025,
  title={Secunda-0.1-RAW},
  author={Yaroster},
  year={2025},
  note={https://huggingface.co/Yaroster/Secunda-0.1-RAW}
}
```

---

## 🪐 From the Cosmos

* [Secunda-0.3-F16-QA](https://huggingface.co/Yaroster/Secunda-0.3-F16-QA) — experimental question-answer variant
* [Secunda-0.3-F16-TEXT](https://huggingface.co/Yaroster/Secunda-0.3-F16-TEXT) — for less structured generation
* [Primétoile](https://yaroster.com) — full VN pipeline

---

⋆°.☾ Secunda-0.1-RAW ☽.°⋆

> ✧ Because every visual novel deserves to begin with a spark of magic ✧

⚠️ This repo contains **only the LoRA adapter weights**. To use the model, download the base `LLaMA 3.1` from Meta (terms apply): [https://ai.meta.com/resources/models-and-libraries/llama-downloads/](https://ai.meta.com/resources/models-and-libraries/llama-downloads/)