                          
       ▄▄▄▄▄    ▄███▄   ▄█▄      ▄      ▄   ██▄   ██       ████▄     ▄███
       █    ▀▄  █▀   ▀  █▀ ▀▄     █      █  █  █  █ █     █   █       ██ 
    ▄   ▀▀▀▀▄   ██▄▄    █   ▀  █   █ ██   █ █   █ █▄▄█     █   █    ▄███▄
     ▀▀▄▄▄▄▀    █▄   ▄▀ █▄  ▄▀ █   █ █ █  █ █  █  █  █     █   █       ▐█ 
               ▀███▀   ▀███▀  █▄ ▄█ █  █ █ ███▀      █     ████▀ ▐█  ▄██▀  
                               ▀▀▀  █   ██         █                          
                                                 ▀                           
   
                      ⋆⋆୨୧˚ THE PRIMÉTOILE ENGINE ˚୨୧⋆｡˚⋆
                  — Visual Novel generation under starlight —

Version	Type	Strengths	Weaknesses	Recommended Use
Secunda-0.1-GGUF / RAW	Instruction	- Most precise - Coherent code - Perfected Modelfile	- Smaller context / limited flexibility	Production / Baseline
Secunda-0.3-F16-QA	QA-based Input	- Acceptable for question-based generation	- Less accurate than 0.1 - Not as coherent	Prototyping (QA mode)
Secunda-0.3-F16-TEXT	Text-to-text	- Flexible for freeform tasks	- Slightly off - Modelfile-dependent	Experimental / Text rewrite
Secunda-0.3-GGUF	GGUF build	- Portable GGUF of 0.3	- Inherits 0.3 weaknesses	Lightweight local testing
Secunda-0.5-RAW	QA Natural	- Best QA understanding - Long-form generation potential	- Inconsistent output length - Some instability	Research / Testing LoRA
Secunda-0.5-GGUF	GGUF build	- Portable, inference-ready version of 0.5	- Shares issues of 0.5	Offline experimentation
Secunda-0.1-RAW	Instruction	- Same base as 0.1-GGUF	- Same as 0.1	Production backup

⋆⁺₊⋆Secunda ☾ 0.3 - F16-TEXT⋆⁺₊⋆

Secunda-0.3-F16-TEXT is the third iteration of the Secunda visual novel generation series, built to transform natural language story prompts into fully structured, high-quality Ren'Py scripts — without relying on explicit instructions or formatting. It was fine-tuned on a carefully curated dataset of fictional .rpy scenes, each paired with its original narrative concept.

This model complements its sibling, Secunda-0.3-F16-QA, which was trained on QA-form data.

🚀 Model Highlights

Instruction-Free Generation: Unlike Secunda-0.1, this version relies only on a natural prompt such as "A detective wakes up in a town where no one remembers him."
Finetuned with QLoRA (FP16): Efficient low-resource finetuning using 16-bit precision on LLaMA 3.1 8B.
Script-Style Outputs: Generates structured .rpy files, often including characters, backgrounds, dialogue, and a return.
More Creative Freedom: Produces more diverse narratives and styles, occasionally multi-scene outputs.

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

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

🔧 Usage

Install required libraries:

pip install -U transformers peft accelerate torch datasets bitsandbytes safetensors

Inference example:

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-0.3-F16-TEXT"

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)

prompt = "A girl wakes up in a town where no one remembers her — except a black cat."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.85, top_p=0.95)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

🧠 Training Details

Base model: meta-llama/Meta-Llama-3.1-8B
Finetuning: QLoRA 4-bit adapters (FP16 runtime)
Data: 800+ handcrafted prompt/script pairs in .jsonl, each with a raw natural language idea and a full .rpy script.
Hardware: NVIDIA RTX 4070

📝 Prompt Style

This model works without instructions. Just write your idea naturally:

"An abandoned café reopens every full moon to serve ghosts their last coffee."

⚠️ Known Issues

May occasionally produce multiple unrelated scripts in one output if token limit is too high.
Less deterministic than instruction-based versions — retries may help.
Not always guaranteed to end with return (you may add it manually).

📌 Tips

Keep prompts between 10–30 words for best coherence.
If the output includes multiple scenes, you can split them manually.
If needed, combine with sanitize_output() logic from Secunda-0.1-RAW to postprocess outputs.

🔒 Privacy

This model is private and intended for research and internal tooling for the Primétoile visual novel engine.

📚 License

Apache 2.0 — For research, testing, and development only.

✨ Credits

Trained and maintained by Yaroster for the Secunda engine inside the Primétoile framework.

Yaroster
/

Secunda-0.3-F16-TEXT