How databrickslabs/dolly works
Dolly is an instruction-following language model, with this repository providing fine-tuning and response-generation code. The principal workflow configures a training run, prepares instruction/response examples and a pretrained model, fine-tunes and saves the model, then loads that output for instruction-based generation. GPU-family configurations and the documented Hugging Face inference workflow are also represented. Edges reflect shown calls, writes, or README statements; unsampled implementation details are not inferred.
Main entry points: generate.py Generation pipeline · train_dolly.py Training notebook · trainer.py Fine-tuning trainer · consts.py Prompt constants
Components 11
Training orchestration
- Training notebook
train_dolly.py
Fine-tuning
- Prompt constants
training/consts.py - Fine-tuning trainer
training/trainer.py
Response generation
- Generation pipeline
training/generate.py
GPU configuration
- GPU configs
config
Other components
- Model user
- Dolly instruction data
- Pretrained model
- Fine-tuned model
- Transformers pipeline
- Generated response
Connections 14
- Model user → Training notebook: configures run
train_dolly.py - Training notebook → GPU configs: selects config
train_dolly.py - Training notebook → Fine-tuning trainer: invokes training
train_dolly.py - Training notebook → Prompt constants: imports settings
train_dolly.py - Fine-tuning trainer → Prompt constants: uses formats
training/trainer.py - Dolly instruction data → Fine-tuning trainer: training examples
README.md - Pretrained model → Fine-tuning trainer: fine-tunes
README.md - Fine-tuning trainer → Fine-tuned model: saves model
training/trainer.py - Training notebook → Generation pipeline: loads and calls
train_dolly.py - Generation pipeline → Prompt constants: uses format
training/generate.py - Model user → Generation pipeline: provides instruction
training/generate.py - Generation pipeline → Fine-tuned model: loads model
- Generation pipeline → Transformers pipeline: subclasses pipeline
training/generate.py - Generation pipeline → Generated response: returns text
training/generate.py