Custom Training

[Ritikavarshney-hub]

Provide fine-tuning options or allow community members to train models on domain-specific spam datasets for improved accuracy

[eccentriccoder01]

Thanks for the suggestion @Ritikavarshney-hub! Since this repo is public and GSSoC25 counts only GitHub PRs, we can’t share Kaggle/Hugging Face credentials. However, you can absolutely enable fine-tuning via PRs. The plan:

• Add a training pipeline (Hugging Face + optional PEFT/LoRA) with configs in configs/
• Add dataset loaders/validators for spam CSVs and stratified splits
• Add an evaluation script (precision/recall/F1, confusion matrix) that saves a JSON/MD report
• Make the app accept a custom Hugging Face model ID (and optional LoRA adapter) for inference
• Add docs (FINE_TUNING.md) explaining how to fine-tune locally/Kaggle using the scripts
• Optionally add a tiny CPU “dry run” in CI to validate the training flow

How Kaggle fits: notebook will be versioned in this repo. You can submit PRs modifying those notebook or the training scripts, and I’ll run them on Kaggle with my credentials and publish resulting models on Hugging Face. That way, all your contributions still count as GitHub PRs.

If you’re interested, please comment which task you want to pick and I’ll create a focused issue with guidance and starter configs.

[Kavlin-Kaur]

@eccentriccoder01 can u please assign this to me!!!
Thank you so much

[yesh00008]

@eccentriccoder01 can u please assign this to me