# Phase 4 - Many-to-One LSTM and Split-Data Hybrids

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

**Download this phase report:** [PDF](phase4_report.pdf) | [Word (.docx)](phase4_report.docx)

## Setup
- Goal: add a many-to-one LSTM baseline across embeddings and re-run the BiLSTM hybrids on
  consistent split data.
- LSTM config: embedding_dim 100, hidden_dim 128, dropout 0.2, max_seq_len 100, Adam lr 1e-2,
  30 epochs, batch 16. Embeddings: Word2Vec, GloVe, FastText.

## Results (many-to-one LSTM, regular training)

| Model | Accuracy | F1 | AUC-ROC |
|-------|----------|------|---------|
| GloVe+LSTM | 0.6779 | 0.5644 | 0.7532 |
| Word2Vec+LSTM | 0.6675 | 0.5892 | 0.7294 |
| FastText+LSTM | 0.6610 | 0.5729 | 0.7208 |

## Findings
- The plain many-to-one LSTM underperforms the BiLSTM hybrids by ~0.07 to 0.10 F1, confirming
  the value of bidirectional context for code-mixed text.
- High specificity but low recall: the LSTM variants lean toward the majority (non-hate) class.

## Files
- Notebooks: [`notebooks/`](https://huggingface.co/pankajbiswas6/hinglish-hate-speech-bilstm/tree/main/phase4/notebooks) - many_to_one_lstm, ManytooneLSTM
- Figures: [`figures/`](https://huggingface.co/pankajbiswas6/hinglish-hate-speech-bilstm/tree/main/phase4/figures) - per-language accuracy curves
- Tables: [`tables/`](https://huggingface.co/pankajbiswas6/hinglish-hate-speech-bilstm/tree/main/phase4/tables) - master_metrics, metrics
- Models: [`models/`](https://huggingface.co/pankajbiswas6/hinglish-hate-speech-bilstm/tree/main/phase4/models) - combined/english/hindi/hinglish BiLSTM