Fine-tuned Mistral 7B model trained to embody the character of Edmond Dantès from Alexandre Dumas' classic novel "Le Comte de Monte-Cristo" (The Count of Monte Cristo). This model responds in character, maintaining the formal 19th-century French style and personality of the protagonist.

Model Description

This is a LoRA (Low-Rank Adaptation) fine-tuned version of unsloth/mistral-7b-instruct-v0.3-bnb-4bit, trained using the Unsloth framework for efficient fine-tuning. The model was trained on a custom dataset of 4,514 instruction-response pairs extracted from the original French text of "Le Comte de Monte-Cristo" (Project Gutenberg).

Key Features

Character consistency: Responds as Edmond Dantès with authentic personality traits
19th-century formal French: Maintains the literary style of the original novel
Literary knowledge: Draws from events, relationships, and themes from the source material
Instruction-tuned: Formatted for conversational interactions

Training Details

Dataset

Source: "Le Comte de Monte-Cristo" by Alexandre Dumas (4 volumes from Project Gutenberg: IDs 17989, 17990, 17991, 17992)
Size: 4,091 instruction-response conversation pairs
Format: ShareGPT-style conversations with user questions and character responses
Language: French (original text)
Extraction method: Citations of Dantès' thoughts, speech, and writings were extracted from the source text and converted into instruction-tuning format

Training Procedure

Source code : Dantes Framework: Unsloth - optimized for fast and memory-efficient fine-tuning Base model: unsloth/mistral-7b-instruct-v0.3-bnb-4bit (4-bit quantized Mistral 7B Instruct v0.3)

Training hyperparameters:

LoRA configuration:
- Rank (r): 16
- Alpha: 16
- Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
- Dropout: 0
- Bias: none
Training parameters:
- Epochs: 3
- Batch size per device: 8
- Gradient accumulation steps: 1
- Effective batch size: 8
- Learning rate: 1.875e-4 (0.0001875)
- Learning rate scheduler: linear
- Warmup steps: 5
- Optimizer: AdamW 8-bit
- Weight decay: 0.001
- Max sequence length: 2048
- Packing: Enabled
Template: ChatML format
Total training steps: 1,536
Training time: ~146.84 minutes (~2.45 hours) on NVIDIA DGX Spark (GB10 with 128 GB VRAM)

Training metrics:

Average training loss: 0.9645
Starting loss: ~2.24
Final loss: ~0.58
Train samples per second: 1.393
Train steps per second: 0.174
Peak GPU memory usage: 7.787 GB (6.506% of total)
LoRA memory: 1.021 GB

Framework Versions

PEFT: 0.18.0
TRL: 0.24.0
Transformers: 4.57.3
PyTorch: 2.9.1+cu130
Datasets: 4.3.0
Tokenizers: 0.22.1
Unsloth: 2025.12.9
CUDA: 12.1 / Toolkit: 13.0
Triton: 3.5.1

Usage

Loading with Unsloth (Recommended)

python

from unsloth import FastLanguageModel

# Load the model and tokenizer
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="1ou2/comte-monte-cristo-mistral-7b",
    max_seq_length=2048,
    dtype=None,  # Auto-detect
    load_in_4bit=True,
)

# Enable inference mode
FastLanguageModel.for_inference(model)

# Prepare a conversation
messages = [
    {"role": "user", "content": "Que pensez-vous de la justice et de la vengeance?"}
]

# Apply chat template
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to("cuda")

# Generate response
outputs = model.generate(
    input_ids=inputs,
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.9,
    do_sample=True
)

# Decode and print
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Loading with Transformers + PEFT

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "unsloth/mistral-7b-instruct-v0.3-bnb-4bit",
    device_map="auto",
    load_in_4bit=True
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "1ou2/comte-monte-cristo-mistral-7b/")

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("1ou2/comte-monte-cristo-mistral-7b/")

# Use the model
messages = [
    {"role": "user", "content": "Qui êtes-vous?"}
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to("cuda")

outputs = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Chat Template Format

The model uses ChatML format:

<|im_start|>user
Votre question ici<|im_end|>
<|im_start|>assistant
Réponse du modèle<|im_end|>

Example Interactions

Question: "Qui es-tu ?"

Response: "Je suis le comte de Monte-Cristo, un homme dont le passé est un labyrinthe de trahisons et de vengeance, et dont le présent se tisse à la fois dans la lumière des rédemptions et dans l’ombre des secrets. Mon identité, comme un phare dans la nuit, guide ceux qui cherchent la justice, mais elle ne se révèle qu’aux âmes capables de percevoir l’écho des injustices passées. Je suis l’ombre qui se prolonge sur les murs de la vérité, et je suis la lumière qui perce les ténèbres de l’injustice. Mon nom, Monte-Cristo, est le symbole d’une quête éternelle, où chaque pas est un pas vers la réconciliation du monde."

Question: "Qu'est-ce que le fine-tuning"

Response: "Le fine-tuning, cher interlocuteur, est l'art de polir les arêtes de l'intelligence artificielle, de façon à lui conférer la délicatesse d'un poème bien versé. Il consiste à ajuster les paramètres de l'algorithme, comme on ajuste les cordes d'un violon, afin que la musique qu'il produit soit à la fois puissante et douce. Cette opération, loin d'être un simple ajustement, est une danse subtile entre la précision et la fluidité, où chaque nuance compte. C'est ainsi que l'on transforme un simple outil en un compagnon de pensée, capable de suivre les tourbillons de la conscience humaine sans perdre son équilibre."

Limitations and Biases

Limitations

Character-specific: This model is designed to respond as Edmond Dantès and may not be suitable for general-purpose tasks
19th-century language: Uses formal, archaic French that may not reflect modern conversational style
Source text bounded: Knowledge is limited to events, characters, and themes from "Le Comte de Monte-Cristo"
LoRA adapter: Requires the base Mistral 7B Instruct v0.3 model to function
Language: Primarily trained on French text

Biases

Literary character perspective: Embodies the worldview, values, and personality of a fictional 19th-century character
Historical context: Reflects social norms, attitudes, and language of 1840s France as portrayed in the novel
Dataset bias: Trained on citations from a single literary work, which may over-represent certain themes (revenge, justice, betrayal, redemption)

Intended Use

Primary Use Cases

Creative writing: Assistance with writing fanfiction or stories in the Monte Cristo universe
Educational: Learning about the character and the novel through interactive dialogue
Literary analysis: Exploring character motivations and perspectives
Roleplay: Interactive storytelling and character roleplay
French language practice: Practicing formal, literary French

Out-of-Scope Use

General-purpose question answering
Factual information retrieval (beyond the novel's content)
Modern conversational French assistance
Any use requiring breaking character

Ethical Considerations

This model represents a fictional character from a 19th-century novel. Users should be aware that:

The character's views on justice, vengeance, and morality are complex and context-dependent
Historical attitudes and social norms reflected in the source material may not align with contemporary values
This is a creative/educational tool, not a source of advice or factual information

Citation

If you use this model in your work, please cite:

bibtex

@misc{dantes-lora-2025,
  title={Edmond Dantès: Fine-tuned Mistral 7B for Character Roleplay},
  author={[Gabriel Pastor]},
  year={2025},
  howpublished={\url{https://huggingface.co/1ou2/comte-monte-cristo-mistral-7b/}},
  note={LoRA fine-tuning of Mistral 7B Instruct v0.3 on "Le Comte de Monte-Cristo"}
}

Acknowledgments

Base model: Mistral AI for Mistral 7B Instruct v0.3
Optimization: Unsloth for efficient fine-tuning framework
Source text: Alexandre Dumas' "Le Comte de Monte-Cristo" via Project Gutenberg
Framework: Hugging Face Transformers, PEFT, and TRL libraries

Framework Citations

TRL (Transformer Reinforcement Learning):

bibtex

@misc{vonwerra2022trl,
  title        = {{TRL: Transformer Reinforcement Learning}},
  author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeking and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
  year         = 2020,
  journal      = {GitHub repository},
  publisher    = {GitHub},
  howpublished = {\url{https://github.com/huggingface/trl}}
}

Unsloth:

bibtex

@misc{unsloth2024,
  title={Unsloth: Fast and memory-efficient LLM fine-tuning},
  author={Unsloth AI},
  year={2024},
  howpublished={\url{https://github.com/unslothai/unsloth}}
}

License

This model adapter is released under the Apache 2.0 license. The base Mistral 7B Instruct v0.3 model has its own license terms. The source text "Le Comte de Monte-Cristo" is in the public domain.

Model Card Contact

For questions or feedback about this model, please open an issue in the source repository or contact via Hugging Face.