| --- |
| license: cc-by-4.0 |
| extra_gated_fields: |
| Full Name: text |
| Affiliation (Organization/University): text |
| Country: country |
| DISCLAIMER The model is released for research purposes only and authors do not take any responsibility for any damage or loss arising due to usage of model or any system/model developed using the model: checkbox |
| datasets: |
| - Exploration-Lab/TechING |
| base_model: |
| - meta-llama/Llama-3.2-11B-Vision-Instruct |
| tags: |
| - Techincal Image Understanding |
| pretty_name: LLama-VL-TUG |
| --- |
| # LLama-VL-TUG (A finetuned model for Technical Image Understanding) |
| This is the official model repo of the paper: |
| > TechING: Towards Real World Technical Image Understanding via VLMs |
|
|
| > **Authors:** Tafazzul Nadeem*, Bhavik Shangari*, Manish Rai, Gagan Raj Gupta, Ashutosh Modi |
|
|
| > **Abstract:** *Professionals working in technical domain typically hand-draw (on whiteboard, paper, etc.) technical |
| > diagrams (e.g., flowcharts, block diagrams, etc.) during discussions; however, if they want to edit these later, |
| > it needs to be drawn from scratch. Modern day VLMs have made tremendous progress in image understanding but they |
| > struggle when it comes to understanding technical diagrams. One way to overcome this problem is to fine-tune on |
| > real world hand-drawn images, but it is not practically possible to generate large number of such images. |
| > In this paper, we introduce a large synthetically generated corpus (reflective of real world images) for training |
| > VLMs and subsequently evaluate VLMs on a smaller corpus of hand-drawn images (with the help of humans). |
| > We introduce several new self-supervision tasks for training and perform extensive experiments with various |
| > baseline models and fine-tune Llama 3.2 11B-instruct model on synthetic images on these tasks to obtain LLama-VL-TUG, |
| > which significantly improves the ROUGE-L performance of Llama 3.2 11B-instruct by 2.14x and achieves the best all-round |
| > performance across all baseline models. On real-world images, human evaluation reveals that we achieve minimum |
| > compilation errors across all baselines in 7 out of 8 diagram types and improve the average F1 score of |
| > Llama 3.2 11B-instruct by 6.97x.* |
|
|
| ## Base Model |
| Base model: meta-llama/Llama-3.2-11B-Vision-Instruct |
| Architecture: Vision-Language Transformer |
| Fine-tuning method: LoRA |
|
|
| ## Training Methodology |
| We fine-tuned Llama3.2-11B-Vision-Instruct using LoRA (image encoder as well as |
| text decoder) on the combination of Primary and Self Supervision tasks (described below) |
| using D1 and D2 corpus of [TechING](https://huggingface.co/datasets/Exploration-Lab/TechING) dataset. |
|
|
| **Primary Tasks** |
| 1. **Image2Code**: Generating corresponding [Mermaid](https://mermaid.js.org/) code for a given image. |
| 2. **Description2Code**: Converting natural language descriptions into Mermaid code. |
| 3. **Image2Description**: Generating Descriptions from technical diagram images. |
| 4. **Image Enhancement via Prompt**: Generating Mermaid code of the updated image, given |
| an image and a natural language enhancement prompt. |
|
|
| **Self Supervision Tasks** |
| 1. **Image Enhancement via Description**: Given an image along with a textual |
| description of the target image, produce code that reflects the enhanced description. |
| 2. **Code Enhancement via Prompt**: Given a Mermaid code along with an enhancement |
| prompt, update the code accordingly. |
| 3. **Code Enhancement via Description**: Given a Mermaid code snippet along with a natural |
| language description of the target image, enhance the code to accurately reflect |
| the changes present in the description. |
| 4. **Positive/Negative Image–Code Pair Q&A**: Predict given image–code pair |
| constitutes a valid match or a mismatch. |
| 5. **Partial Match Image–Code Pair Q&A**: Identify partial matches between incomplete |
| and complete image-code pairs. |
|
|
| ## Hyperparameters Details |
| per_device_train_batch_size: 1 |
| gradient_accumulation_steps: 1 |
| learning_rate: 2e-5 |
| weight_decay: 0.05 |
| num_train_epochs: 2 |
| lr_scheduler_type: cosine |
| warmup_ratio: 0.2 |
| bf16: True |
| lora_rank: 32 |
| lora_alpha: 16 |
| target_modules: QKV |
| lora_dropout: 0.2 |
| use_rslora: True |
| ## Evaluation Results |
| The radar charts present ROUGE-L performance across the |
| three primary tasks on the D1 test set, comparing LLama-VL-TUG |
| against baselines of comparable model size. Detailed results are |
| provided in our paper, [TechING: Towards Real World Technical Image Understanding via VLMs](https://arxiv.org/abs/2601.18238). |
| <img src="evaluation_results.png"> |
|
|
| ## Loading the Model |
| To load the model using the huggingface: |
| ```python |
| from transformers import MllamaForConditionalGeneration |
| from peft import LoraConfig, get_peft_model, PeftModel |
| import torch |
| |
| base_model_id = "meta-llama/Llama-3.2-11B-Vision-Instruct" |
| base_model = MllamaForConditionalGeneration.from_pretrained( |
| base_model_id, |
| torch_dtype=torch.bfloat16, |
| device_map="auto" |
| ) |
| peft_model_repo = "Exploration-Lab/LLama-VL-TUG" |
| model = PeftModel.from_pretrained(base_model, peft_model_repo) |
| ``` |
| ## Citation |
|
|
| [**TechING: Towards Real World Technical Image Understanding via VLMs**](https://2026.eacl.org/), In the 19th Conference of the |
| European Chapter of the Association for Computational Linguistics (EACL) to be held in Rabat, Morocco, from March 24–29, 2026. |
| ``` |
| @misc{nadeem2026techingrealworldtechnical, |
| title={TechING: Towards Real World Technical Image Understanding via VLMs}, |
| author={Tafazzul Nadeem and Bhavik Shangari and Manish Rai and Gagan Raj Gupta and Ashutosh Modi}, |
| year={2026}, |
| eprint={2601.18238}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.CL}, |
| url={https://arxiv.org/abs/2601.18238}, |
| } |
| ``` |
|
|
| ## License |
| [](https://creativecommons.org/licenses/by-nc/4.0/) |
| **TechING** follows [CC-BY-NC](CC-BY-NC) license. Thus, users can share and adapt the dataset/codebase if they give credit to the authors and do not use the dataset/codebase for any commercial purposes. |
|
|
| \*Equal Contribution |