RMunshi/vlm-layout-master

Model Card for vlm-student-thesis

This model is a fine-tuned version of unsloth/llama-3.2-11b-vision-instruct-bnb-4bit. It has been trained using TRL.

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="RMunshi/vlm-student-thesis", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])

Training procedure

Training Metrics & Thesis Documentation

This model was fine-tuned for a Master's Thesis on PDF Layout Distillation. The training process followed a "Student-Teacher" architecture where the knowledge from GPT-4o-V was distilled into this Llama-3.2-11B-Vision instance.

📊 Training Curves

The primary training metrics are tracked in real-time on Weights & Biases: **View Live Training Curves (W&B)**

Key results:

• —Final Loss: ~0.06 - 0.08 (Strong convergence)
• —Training Epochs: 2.22
• —Hardware: NVIDIA RTX 3090 (24GB VRAM)

📂 Reproducibility

The training script used to generate this model is available in the repository as train_student.py.

Framework versions

• —TRL: 0.24.0
• —Transformers: 4.57.6
• —Pytorch: 2.5.1
• —Datasets: 4.3.0
• —Tokenizers: 0.22.2

Citations

Cite TRL as:

@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}}
}