This repository contains my team's solution for the Mid Prep Problem Statement of InterIIT TechMeet 13.0, "Adobe Behaviour Simulation Challenge" by Adobe.
- 1. Image_Captioning.ipynb: In this notebook we have downloaded the media files and generate caption for them using BLIP model.
- 2. 100K_Data_Generation.ipynb: In this notebook we have reduced the length of the actual dataset (300K) and only consider the first 100K samples.
- 3. Data_Cleaning.ipynb: In this notebook we have combined the tokenized content text with the generated captions along with their embeddings, and then splitted this into train and test datasets.
- Tokenization.ipynb: In this notebook we have tokenized the content (captions) of the given dataset using the DistilBERT tokenizer.
- Creating_Prompts.ipynb: In this notebook we have written prompt to convert the data into more organized format for the bucketing process.
- Bucketing.ipynb: In this notebook we have created 7 buckets (classes) and trained the DistilBERT model for this classification task.
- FineTune.ipynb: In this notebook we have Fine Tuned the Bloom 7b and Qwen 1.5b models using LoRA and PEFT techniques.
- Inference.ipynb: In this notebook, using the fine tuned models we have performed predictions on the first 500 samples.
- Test_Metrics.ipynb: In this notebook we have used BLEU (1-4), ROUGE (1,2,L) and CIDEr metrics to evaluate the model's predictions with the original captions.