Petite Elle L'Aime 3

Language Understanding Tasks

• Sports Comprehension: petite-elle-L-aime-3-sft achieves 57.6% accuracy, outperforming SmolLM3 by 3.6 percentage points
• Causal Judgment: Strong performance at 57.2% accuracy, slightly better than SmolLM3
• Hyperbate (Adjective Ordering): petite-elle-L-aime-3-sft shows superior performance at 53.6% vs 52.8%

Mathematical Reasoning

• Prealgebra: petite-elle-L-aime-3-sft's strongest math performance at 48.7% accuracy, outperforming SmolLM3 by 0.9 percentage points
• Geometry: Matches SmolLM3's performance exactly at 11.7%

Question Answering

• GPQA Diamond: petite-elle-L-aime-3-sft performs slightly better at 34.0% vs 33.5%
• Murder Mysteries: petite-elle-L-aime-3-sft shows competitive performance at 50.4% vs 49.6%

Object Tracking

• 5-Object Tracking: petite-elle-L-aime-3-sft performs better at 18.0% vs 16.0% accuracy
• This suggests petite-elle-L-aime-3-sft may have better spatial reasoning for moderate complexity tasks

petite-elle-L-aime-3-sft's Performance Profile

Areas of Excellence

1. Sports Domain Knowledge: petite-elle-L-aime-3-sft demonstrates superior understanding of sports-related content
2. Causal Reasoning: Strong performance in understanding cause-and-effect relationships
3. Adjective Ordering: Better grasp of linguistic rules for adjective placement
4. Basic Mathematical Operations: Competitive performance in prealgebra tasks
5. Spatial Reasoning: Better performance in moderate complexity object tracking

Competitive Performance

• MMLU: Nearly identical performance (50.55% vs 50.60%)
• Instruction Following: Similar performance on strict evaluation metrics
• Geometric Recognition: Matches SmolLM3's performance exactly

Model Architecture

This is a fine-tuned version of the SmolLM3-3B model with the following specifications:

• Base Model: SmolLM3-3B
• Parameters: ~3B
• Context Length: 128k
• Languages: English, French , Italian , Portugese , Chinese , Arabic
• Architecture: Transformer-based causal language model

Performance

The model provides:

• Text Generation: High-quality text generation capabilities
• Conversation: Natural conversation abilities
• Multilingual: Support for English and French
• Quantized Versions: Optimized for different deployment scenarios

Limitations

1. Context Length: Limited by the model's maximum sequence length
2. Bias: May inherit biases from the training data
3. Factual Accuracy: May generate incorrect or outdated information
4. Safety: Should be used responsibly with appropriate safeguards
5. Quantization: Quantized versions may have slightly reduced accuracy

Usage

Requirements

pip install torch transformers accelerate
pip install torchao  # For quantized models

Hardware Requirements

• Main Model: GPU with 8GB+ VRAM recommended
• int4 Model: CPU deployment possible

Quick start

from transformers import pipeline

question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
generator = pipeline("text-generation", model="None", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])

Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load the main model
model = AutoModelForCausalLM.from_pretrained(
    "Tonic/petite-elle-L-aime-3-sft",
    device_map="auto",
    torch_dtype=torch.bfloat16
)
tokenizer = AutoTokenizer.from_pretrained("Tonic/petite-elle-L-aime-3-sft")

# Generate text
input_text = "What are we having for dinner?"
input_ids = tokenizer(input_text, return_tensors="pt").to(model.device.type)
output = model.generate(**input_ids, max_new_tokens=50)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Quantized Models

This repository also includes quantized versions of the model for improved efficiency:

int4 Weight-Only Quantization (CPU Optimized)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load int4 quantized model (CPU optimized)
model = AutoModelForCausalLM.from_pretrained(
    "Tonic/petite-elle-L-aime-3-sft/int4",
    device_map="cpu",
    torch_dtype=torch.bfloat16
)
tokenizer = AutoTokenizer.from_pretrained("Tonic/petite-elle-L-aime-3-sft/int4")

Quantization Benefits

• int4 (CPU): ~50% memory reduction, significantly faster inference with some accuracy trade-off

Training Information

Training Configuration

• Base Model: HuggingFaceTB/SmolLM3-3B
• Dataset: legmlai/openhermes-fr
• Training Config: SFT
• Trainer Type: TRL
• Dataset Sample Size: 800K

Training Parameters

• Batch Size: 12
• Gradient Accumulation: 12
• Learning Rate: 4e-6
• Max Epochs: 1.6
• Sequence Length: 16384

Training Infrastructure

• Hardware: A100 SXM
• Monitoring: Track Tonic
• Experiment: exp_20250727_172526

Monitoring and Tracking

This model was trained with custom monitoring:

• Track Tonic Space: [Track Tonic]https://huggingface.co/spaces/Tonic/track-tonic
• Experiment: exp_20250727_172526
• Dataset Repository: trackio-experiments
• Training Logs: Available in the experiment data or on the Track Tonic Spaces

Training Data

The model was fine-tuned on:

• Dataset: legmlai/openhermes-fr
• Size: 800K
• Format: DPO
• Languages: French

Evaluation

The model was evaluated using:

• Metrics: le-leadboard/OpenLLMFrenchLeaderboard
• Monitoring: Real-time tracking via Track Tonic
• Validation: Regular validation during training

Framework versions

• TRL: 0.19.1
• Transformers: 4.54.0
• Pytorch: 2.7.1+cu118
• Datasets: 4.0.0
• Tokenizers: 0.21.2

Acknowledgments

• Base Model: SmolLM3-3B by HuggingFaceTB
• Training Framework: PyTorch, Transformers, PEFT
• Monitoring: Trackio integration
• Quantization: torchao library

Cite

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

@misc{Petite_Elle_L_Aime_3_SFT,
  title={{Petite Elle L'Aime 3}},
  author={{Joseph Pollack}},
  year={2025},
  url={https://huggingface.co/Tonic/petite-elle-L-aime-3-sft}
}

Citations

@misc{vonwerra2022trl,
	title        = {{TRL: Transformer Reinforcement Learning}},
	author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching 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}}
}

License

This model is licensed under the Apache 2.0 License.

Changelog

• v1.0.0: Initial release with fine-tuned model
• v1.1.0: Added quantized versions (int4)

Repository Structure

petite-elle-L-aime-3-sft/
├── .gitattributes                         # 1.96 kB, Git configuration for file handling
├── README.md                              # 7.38 kB, Project overview and instructions (updated 7 minutes ago)
├── chat_template.jinja                    # 5.6 kB, Chat template
├── config.json                            # 1.92 kB, Model configuration
├── generation_config.json                 # 177 Bytes, Generation settings
├── model-00001-of-00002.safetensors       # 4.97 GB, Model weights (part 1)
├── model-00002-of-00002.safetensors       # 1.18 GB, Model weights (part 2)
├── model.safetensors.index.json           # 26.9 kB, Model index
├── pytorch_model.bin                      # Main model weights
├── special_tokens_map.json                # 289 Bytes, Special tokens
├── tokenizer.json                         # 17.2 MB, Tokenizer data
├── tokenizer_config.json                  # 50.4 kB, Tokenizer configuration
├── train_results.json                     # 182 Bytes, Training results
├── training_args.bin                      # 6.16 kB, Training arguments
├── training_results/                      # Training results directory
├── checkpoint-4000/                       # Training checkpoint at step 4000
├── checkpoint-5000/                       # Training checkpoint at step 5000
├── checkpoint-6000/                       # Training checkpoint at step 6000
├── checkpoint-7000/                       # Training checkpoint at step 7000
├── checkpoint-8000/                       # Training checkpoint at step 8000
│   ├── chat_template.jinja                # 5.6 kB, Chat template
│   ├── config.json                        # 1.92 kB, Checkpoint configuration
│   ├── generation_config.json             # 177 Bytes, Generation settings
│   ├── model-00001-of-00002.safetensors   # 4.97 GB, Model weights (part 1)
│   ├── model-00002-of-00002.safetensors   # 1.18 GB, Model weights (part 2)
│   ├── model.safetensors.index.json       # 26.9 kB, Model index
│   ├── optimizer.pt                       # 12.3 GB, Optimizer state
│   ├── rng_state.pth                      # 14.6 kB, Random number generator state
│   ├── scheduler.pt                       # 1.47 kB, Scheduler state
│   ├── special_tokens_map.json            # 289 Bytes, Special tokens
│   ├── tokenizer.json                     # 17.2 MB, Tokenizer data
│   ├── tokenizer_config.json              # 50.4 kB, Tokenizer configuration
│   ├── trainer_state.json                 # 84.5 kB, Trainer state (uploaded ~4 hours ago)
│   └── training_args.bin                  # Training arguments
└── int4/                                  # Quantized model for CPU
    ├── README.md                          # 1.7 kB, Quantized model documentation (updated ~1 hour ago)
    ├── config.json                        # 2.67 vikB, Quantized model configuration
    ├── generation_config.json             # 177 Bytes, Generation settings
    ├── pytorch_model.bin                  # 2.63 GB, Quantized model weights
    ├── special_tokens_map.json            # 289 Bytes, Special tokens
    ├── tokenizer.json                     # 17.2 MB, Tokenizer data
    └── tokenizer_config.json              # Tokenizer configuration