Ym512/amazigh-odyssey-gm

Amazigh Odyssey — Game Master (LoRA)

A fine-tuned LoRA adapter for Mistral Nemo 12B that acts as the Game Master of Amazigh Odyssey, a narrative RPG set in mythical ancient North Africa (Numidia).

The model generates immersive narration in French enriched with Kabyle vocabulary, Tifinagh script, Amazigh mythology, and structured game mechanics (choices, effects, combat).

Model Details

Base model	mistralai/Mistral-Nemo-Instruct-2407 (12B)
Adapter type	LoRA (PEFT)
LoRA rank (r)	32
LoRA alpha	64
LoRA dropout	0.05
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Trainable params	114M / 12.2B (0.92%)
Adapter size	~456 MB

Training

Dataset	286 custom samples (139 KB JSONL)
Format	Mistral SFT chat format ({"messages": [...]})
Categories	Narration, Kabyle dialogue, combat, Tifinagh magic, proverbs, cultural lore, exploration
Epochs	3
Batch size	2 (effective 16 with grad_accum=8)
Learning rate	2e-4 (cosine schedule, 10% warmup)
Precision	bf16
Hardware	NVIDIA L40S 48GB
Training time	2 min 54 sec

Metrics

Metric	Value
Train loss	3.54 → 0.47
Eval loss	1.019
Token accuracy	90%
Peak VRAM	35.3 GB

Dataset

The training data covers 7 categories of game master interactions:

• Narration — Scene descriptions for locations (Cirta, Tassili caves, Atlas mountains, Sahara)
• Kabyle dialogue — NPC conversations with Kabyle phrases and Tifinagh script (Dihya, Tin Hinan, Jugurtha, Massinissa)
• Combat — Encounters with mythical creatures (djinns, stone guardians, cursed bulls)
• Tifinagh magic — Spell casting using glyph combinations (fire+air, water+earth, etc.)
• Proverbs — Integration of authentic Kabyle proverbs into gameplay
• Cultural lore — Amazigh pantheon (Ammon, Tanit, Anzar, Ifri, Gurzil, Tafukt)
• Exploration — Discovery events, environmental storytelling

Usage

With transformers + PEFT

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained(
    "mistralai/Mistral-Nemo-Instruct-2407",
    device_map="auto",
    torch_dtype="auto",
)
model = PeftModel.from_pretrained(base, "Ym512/amazigh-odyssey-gm")
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-Nemo-Instruct-2407")

messages = [
    {"role": "user", "content": "Le joueur entre dans les Grottes du Tassili pour la première fois. Décris la scène."}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=512, temperature=0.8, do_sample=True)
print(tokenizer.decode(output[0], skip_special_tokens=True))

With vLLM (recommended for serving)

pip install vllm
vllm serve mistralai/Mistral-Nemo-Instruct-2407 \
    --enable-lora \
    --lora-modules amazigh-gm=Ym512/amazigh-odyssey-gm \
    --port 8000

With the Amazigh Odyssey backend

# In .env:
USE_FINETUNED_GM=true
FINETUNED_GM_URL=http://localhost:8000

# Start serving, then start the game backend
python -m finetune.serve_model
python -m uvicorn backend.server:app --port 8765

Example Output

Prompt: "Le joueur rencontre Dihya pour la première fois."

Response:

Une femme se tient au sommet du rocher, son armure de cuir gravée de glyphes tifinagh, ses yeux d'ambre te transpercent.

"ⵜⵉⵍⴻⵍⵍⵉ... Tilelli. Liberté. C'est le premier mot qu'on apprend dans ces montagnes."

Elle croise les bras.

"Tu portes un glyphe que tu ne comprends pas encore. Qui t'a donné ce pouvoir, étranger ?"

Project

Part of Amazigh Odyssey — Numidia Mythica, a 3D RPG powered by Mistral AI multi-agents. Built for the Mistral AI hackathon.

• Game backend: FastAPI + WebSocket with 4 AI agents (Game Master, Narrator, Combat, Dialogue)
• 3D world: Godot 4 with AI-connected NPCs
• Cultural data: 289 entries of Kabyle vocabulary, proverbs, mythology, songs

License

Apache 2.0 (same as base model)