# Phase 5 - GloVe+BiLSTM Fine-Tune and SHAP Explainability

[Back to model card](../README.md) | [Dataset repository](https://huggingface.co/datasets/pankajbiswas6/prism-hinglish-hate-speech)

**Download this phase report:** [PDF](phase5_report.pdf) | [Word (.docx)](phase5_report.docx)

## Setup
- Goal: fine-tune the best hybrid (GloVe+BiLSTM) with a HHE (Hindi-Hinglish-English) regime
  and add explainability via SHAP.
- Per-language best models saved: english_best, hindi_best, hinglish_best, full_dataset_best.

## Findings
- SHAP confirms the model keys on profanity/abuse tokens for the hate class but is misled by
  transliterated slurs and code-switch boundaries (visible in the false-negative force plots).
- False positives cluster around aggressive-but-non-hateful phrasing; false negatives around
  obfuscated or romanized Hindi abuse, a concrete target for future preprocessing.

## Files
- Figures: [`figures/`](https://huggingface.co/pankajbiswas6/hinglish-hate-speech-bilstm/tree/main/phase5/figures) - per-language confusion matrix, ROC, PR, F1, train curves
- SHAP: [`shap/`](https://huggingface.co/pankajbiswas6/hinglish-hate-speech-bilstm/tree/main/phase5/shap) - false-positive and false-negative force plots plus waterfall samples
  (representative subset; the full set of ~1,390 plots was generated during training)
- Tables: [`tables/`](https://huggingface.co/pankajbiswas6/hinglish-hate-speech-bilstm/tree/main/phase5/tables) - final_results, fp_fn_samples
- Models: [`models/`](https://huggingface.co/pankajbiswas6/hinglish-hate-speech-bilstm/tree/main/phase5/models) - english/hindi/hinglish/full_dataset best (.h5)