Fine-Tune Tamil Meme Classifier
Budget: ₹12,500 – ₹37,500 INR
I have already put together a complete, end-to-end pipeline that scrapes Tamil political memes from Facebook, Twitter, Instagram and a handful of public websites, strips out the noise in both the images and the Tamil text, stores everything neatly, and routes it through native-speaker annotations.
Where I need an extra pair of expert hands is the model-building and fine-tuning stage. So far I have experimented with:
• CLIP (ViT-B/32) in a few-shot setup with Chain-of-Thought prompting
• BLIP-2 using a two-phase fine-tuning strategy
• A MuRIL encoder fused with ResNet-50 for joint text-image understanding
These baselines reach 81.1 % at the coarse Level-1 sentiment split (Support/Praise vs. Troll/Oppose) and 58.2 % at the finer Level-2 categories on a 244-item held-out set. I want to push those numbers higher and tighten the confusion at Level-2 while keeping the model lightweight enough for practical inference.
You will work with both images and their corresponding Tamil captions. Feel free to bring in any advanced multimodal tricks—better prompt engineering, adapter layers, parameter-efficient tuning, vision encoder swaps, contrastive pre-training, you name it—as long as the final solution is reproducible in PyTorch, integrates cleanly with my current data manager, and shows clear gains on accuracy and ROC-AUC.
Deliverables I expect
1. Updated training code or notebooks with commentary
2. The fine-tuned model weights (or adapters) ready for inference
3. A short evaluation report comparing your approach to the existing baselines, including new confusion matrices and ROC curves
Acceptance criteria
• Minimum +3 % absolute improvement at Level-2 accuracy on the same test set
• All code runs on CUDA 11.7 with a single 24 GB GPU without modification
If this sounds like your kind of challenge, let’s get started.
Where I need an extra pair of expert hands is the model-building and fine-tuning stage. So far I have experimented with:
• CLIP (ViT-B/32) in a few-shot setup with Chain-of-Thought prompting
• BLIP-2 using a two-phase fine-tuning strategy
• A MuRIL encoder fused with ResNet-50 for joint text-image understanding
These baselines reach 81.1 % at the coarse Level-1 sentiment split (Support/Praise vs. Troll/Oppose) and 58.2 % at the finer Level-2 categories on a 244-item held-out set. I want to push those numbers higher and tighten the confusion at Level-2 while keeping the model lightweight enough for practical inference.
You will work with both images and their corresponding Tamil captions. Feel free to bring in any advanced multimodal tricks—better prompt engineering, adapter layers, parameter-efficient tuning, vision encoder swaps, contrastive pre-training, you name it—as long as the final solution is reproducible in PyTorch, integrates cleanly with my current data manager, and shows clear gains on accuracy and ROC-AUC.
Deliverables I expect
1. Updated training code or notebooks with commentary
2. The fine-tuned model weights (or adapters) ready for inference
3. A short evaluation report comparing your approach to the existing baselines, including new confusion matrices and ROC curves
Acceptance criteria
• Minimum +3 % absolute improvement at Level-2 accuracy on the same test set
• All code runs on CUDA 11.7 with a single 24 GB GPU without modification
If this sounds like your kind of challenge, let’s get started.