by Aleksei Vdonin
Faculty of Computer Science, HSE University, Moscow, 2025
A self-hosted micro-service that recognises and summarises HSE corporate documents (PDF, DOC/DOCX, TXT) and is ready to be plugged into the mobile application HSE App X.
The pipeline can run end-to-end on CPU-only servers inside the HSE perimeter, keeping sensitive documents private.
| Stage | Description |
|---|---|
| 1. Crawling & OCR | Parses >2 000 public documents from hse.ru and runs Tesseract OCR where needed. |
| 2. Dataset Builder | Splits texts into 1 000-token chunks with 100-token overlap, then queries OpenAI o3-mini to create reference summaries. |
| 3. Teacher | Fine-tunes d0rj/ru-mbart-large-summ to reach state-of-the-art quality on legal prose. |
| 4. Students | Distils the teacher into compact BiLSTM and BiLSTM + MHA Seq2Seq models for CPU inference. |
| 5. Streamlit UI | Lets users upload a file or paste text and instantly compare outputs of all four models. |
All components are container-friendly and require only Python 3.11 + PyPI packages.
- Document-scale abstractive summarization (arbitrary length, sliding window).
- Soft-label knowledge distillation for ~5× faster CPU inference.
- End-to-end automation: call → OCR → chunk → target → summary.
- Lightweight deployment (no GPU, < 200 MB student weights).
- Interactive web demo with timing & progress bars.
| Stage | Output | Size |
|---|---|---|
| Crawl portal | data/data_raw/ |
2 059 docs |
| OCR & cleaning | data/rproc_data/ |
2 059 UTF-8 texts |
| Chunking | 8 000 + fragments | ≈ 860 tokens each |
| Final JSONL | train_smart.jsonl |
~67 MB (text, summary pairs) |
| Stage | Entry-point | Epochs | GPU-hours |
|---|---|---|---|
| Teacher fine-tune | train/mBART.ipynb |
7 | 27 |
| LSTM student | distill/distill_train.ipynb |
50 | 30 |
| LSTM + MHA student | distill/distill_train_mha.ipynb |
16 | 20 |
Experiments are logged with Weights & Biases; reproducible configs are present in respective folders.
| Model | Params | BERTScore F1 | CPU latency (500 w) | Speed-up |
|---|---|---|---|---|
| Teacher (mBART-large) | 380 M | 0.76 | 11.2 s | 1× |
| Student (LSTM + MHA) | 47 M | 0.69 | 4.66 s | 2.4× |
| Student (LSTM) | 46 M | 0.68 | 2.23 s | 5.0× |
The Streamlit interface supports:
- File upload (TXT, DOC/DOCX, PDF with on-the-fly OCR).
- Real-time comparison of all models with elapsed-time badges.
- Progress bars for OCR & long-text chunking.
A live instance runs at http://158.160.61.64:8501/ (internal HSE network).
- Hybrid extractive + abstractive summarization for better factual coverage.
- Dynamic model selector (fast vs. accurate) based on document length / SLA.
- ONNX export of student models for mobile devices.
Released under the MIT License – see LICENSE for full text.