How duyhai020703/danlp works
DANLP is a notebook-led workflow for adapting BLOOM-560M to Vietnamese conversational QA and serving the resulting model. Its documented path runs from JSON corpus preprocessing and prompt/token preparation through LoRA fine-tuning and adapter export, then through model integration and controlled decoding to a Flask chat API, web UI, and optional public tunnel. The repository includes the notebook and datasets, but no source excerpts were provided; implementation links below reflect README-documented architecture, not inspected notebook code.
Main entry points: NLP_DA .ipynb Causal LM training · train.json Training corpus · test.json Evaluation corpus
Components 16
Data preparation
- Training corpus
Data_DA/train.json - Evaluation corpus
Data_DA/test.json - Clean and normalize
NLP_DA .ipynb - Prompt formatting
NLP_DA .ipynb - BLOOM tokenization
NLP_DA .ipynb
Model training
- LoRA adapter setup
NLP_DA .ipynb - Causal LM training
NLP_DA .ipynb - Training metrics
NLP_DA .ipynb - Exported adapter
NLP_DA .ipynb
Inference and serving
- PEFT integration
NLP_DA .ipynb - Controlled decoding
NLP_DA .ipynb - Flask chat API
NLP_DA .ipynb - Web interface
NLP_DA .ipynb - Ngrok tunnel
NLP_DA .ipynb
Other components
- BLOOM-560M
- Chat user
Connections 16
- Training corpus → Clean and normalize: provides records
README.md - Clean and normalize → Prompt formatting: cleans text
README.md - Prompt formatting → BLOOM tokenization: formats for
README.md - Evaluation corpus → BLOOM tokenization: provides evaluation data
README.md - BLOOM-560M → LoRA adapter setup: receives adapter
README.md - LoRA adapter setup → Causal LM training: configures
README.md - BLOOM tokenization → Causal LM training: supplies tokens
README.md - Causal LM training → Training metrics: logs training
README.md - Causal LM training → Exported adapter: exports weights
README.md - Exported adapter → PEFT integration: loads into
README.md - BLOOM-560M → PEFT integration: provides base weights
README.md - PEFT integration → Controlled decoding: supplies model
README.md - Controlled decoding → Flask chat API: serves responses
README.md - Flask chat API → Web interface: serves interface
README.md - Flask chat API → Ngrok tunnel: exposes endpoint
README.md - Chat user → Web interface: submits chat
README.md