# Phase 2 - Hybrid Baselines (Monolingual)

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

**Download this phase report:** [PDF](phase2_report.pdf) | [Word (.docx)](phase2_report.docx)

## Setup
- Goal: build the first hybrid models (pretrained embeddings feeding a BiLSTM) and measure
  them per language (monolingual).
- Models: FastText+BiLSTM, GloVe+BiLSTM, Word2Vec+BiLSTM.
- Config: embedding_dim 300, hidden_dim 256, dropout 0.5, max_seq_len 128, Adam lr 1e-3,
  10 epochs, batch 32.

## Findings
- BiLSTM over pretrained embeddings clearly beats plain LSTM baselines on code-mixed data.
- GloVe and FastText are competitive; Word2Vec trails slightly on the harder Hinglish subset.
- Monolingual training transfers poorly across languages, motivating the language-wise
  strategy in Phase 3.

## Files
- Notebooks: [`notebooks/`](https://huggingface.co/pankajbiswas6/hinglish-hate-speech-bilstm/tree/main/phase2/notebooks) - fasttext_bilstm_monolingual
- Figures: [`figures/`](https://huggingface.co/pankajbiswas6/hinglish-hate-speech-bilstm/tree/main/phase2/figures) - confusion matrix, ROC, train/val accuracy and loss
- Tables: [`tables/metrics_table.csv`](tables/metrics_table.csv)