How bilalahmed0060/deep-nlp-bertified works
This repository is a coursework NLP modeling and evaluation project: it trains and evaluates BERT-based classifiers and similarity models across sentiment, paraphrase, semantic similarity, and paraphrase-type tasks, alongside BART paraphrase generation and detection. The principal workflow is task data → tokenized batches → task model → evaluation/predictions. The README documents QQP variants including contrastive, SBERT-style, and stemming approaches; other documented task branches are retained. Sampled excerpts are partial, so uncertain wiring between specialized training scripts and shared utilities is omitted.
Main entry points: datasets.py Batch datasets · multitask_classifier.py Multitask BERT · multitask_classifier_qqp.py QQP variants · multitask_classifier_sts.py STS variants
Components 22
Task data
- SST data
data/sst-sentiment-train.csv - QQP data
data/quora-paraphrase-train.csv - STS data
data/sts-similarity-train.csv - ETPC data
data/etpc-paraphrase-train.csv - Batch datasets
datasets.py
BERT tasks
- BERT tokenizer
tokenizer.py - Multitask BERT
multitask_classifier.py - QQP variants
multitask_classifier_qqp.py - STS variants
multitask_classifier_sts.py - SST fine-tuning
sst_finetune_gradual_unfreezing.py - SST R-Drop
sst_finetune_rdrop.py - BERT encoder
bert.py
BART tasks
- Paraphrase generation
bart_generation.py - ETPC detection
finetune_bart_detection_asl.py - Detection variant
finetune_bart_detection-dbloss.py - ASL tuning
asl_hparam_tuner.py
Evaluation and outputs
- Task evaluation
evaluation.py - Prediction files
predictions/bert - BART predictions
predictions/bart - ASL metrics
metrics/asl_metrics.csv
Other components
- Researcher
- Pretrained model source
Connections 31
- Researcher → Multitask BERT: starts training
- Researcher → QQP variants: selects approach
- Researcher → STS variants: starts training
- Researcher → Paraphrase generation: starts generation
- Researcher → ETPC detection: starts detection
- SST data → Batch datasets: loads examples
- QQP data → Batch datasets: loads examples
- STS data → Batch datasets: loads examples
- ETPC data → Batch datasets: loads examples
- Batch datasets → BERT tokenizer: tokenizes batches
- Multitask BERT → Batch datasets: loads task data
- Multitask BERT → BERT encoder: encodes sentences
- QQP variants → Batch datasets: loads task data
- QQP variants → BERT encoder: encodes pairs
- STS variants → Batch datasets: loads task data
- STS variants → BERT encoder: encodes pairs
- SST fine-tuning → BERT encoder: fine-tunes encoder
- SST R-Drop → BERT encoder: fine-tunes encoder
- Multitask BERT → Task evaluation: evaluates model
- QQP variants → Task evaluation: evaluates model
- STS variants → Task evaluation: evaluates model
- Task evaluation → Prediction files: returns predictions
- Paraphrase generation → BART predictions: saves generations
- ETPC detection → ETPC data: loads labels
- ASL tuning → ETPC data: loads training data
- ASL tuning → ASL metrics: records metrics
- Multitask BERT → Pretrained model source: loads checkpoint
- QQP variants → Pretrained model source: loads checkpoint
- STS variants → Pretrained model source: loads checkpoint
- Paraphrase generation → Pretrained model source: loads checkpoint
- ETPC detection → Pretrained model source: loads checkpoint