An archive of completed machine-learning experiments, currently centered on an IMDb sentiment-classification study with traditional baselines, DistilBERT, data-scale experiments, and exploratory error analysis.
Quick Start · At a Glance · Results Snapshot · Featured Project · Repository Structure · Scope and Limitations
| Goal | Open |
|---|---|
| Understand the study | IMDb project README |
| Inspect the experiment workflow | IMDb notebook |
| Review archived metrics and figures | Results directory |
| Check limitations and proposed work | Project limitations · Future work |
The IMDb dataset is not redistributed. Rerunning the notebook requires a separately downloaded dataset and a compatible local Python/CUDA environment.
| Item | Archived study |
|---|---|
| Task | Binary sentiment classification |
| Dataset | IMDb Large Movie Review Dataset — 25,000 train / 25,000 test reviews |
| Models | Multinomial Naive Bayes · Logistic Regression · DistilBERT |
| Best full-data F1 | 0.914496 — DistilBERT |
| Training-data study | 20% · 50% · 100% |
| Error pool | 5,465 test reviews misclassified by at least one evaluated model |
| Status | Completed and archived |
Full training-data setting:
| Model | F1 |
|---|---|
| TF-IDF + Naive Bayes | 0.846688 |
| TF-IDF + Logistic Regression | 0.883284 |
| DistilBERT | 0.914496 |
These values are read from the archived formal result CSV. Timing values are hardware-dependent: the traditional models ran on CPU and DistilBERT used CUDA acceleration.
A completed comparison of sparse traditional baselines and DistilBERT, covering predictive performance, computational cost, training-data scale, and exploratory error patterns. See the complete project README for research questions, model settings, full results, error-analysis boundaries, reproducibility notes, and proposed extensions.
Machine-Learning/
├── IMDb-Sentiment-Analysis/
│ ├── notebooks/ archived experiment workflow
│ ├── results/ exported metrics, tables, and figures
│ ├── docs/assets/ project-level presentation asset
│ └── README.md complete project record
├── docs/assets/ archive-level presentation assets
├── .gitignore
└── README.md
Python · Jupyter Notebook · pandas · NumPy · scikit-learn · PyTorch · Hugging Face Transformers · Matplotlib
- This is a completed research archive, not an actively developed software product.
- The IMDb dataset is not redistributed in this repository.
- The 5,465-review error pool is the union of reviews misclassified by at least one model; it is not a count of DistilBERT-only errors.
- The 210 representative error examples use a balanced inspection design and do not estimate the natural category distribution of the full error pool.
- Rule-based error categories are exploratory, and the archived study has no multi-seed statistical analysis.
- Timing comparisons are practical hardware-dependent measurements, not hardware-neutral benchmarks.
- Difficulty taxonomies, cost-adaptive routing, selective prediction, and other extensions remain proposed and were not implemented.