https://github.com/eleutherai/features-across-time
Understanding how features learned by neural networks evolve throughout training
https://github.com/eleutherai/features-across-time
Last synced: about 1 year ago
JSON representation
Understanding how features learned by neural networks evolve throughout training
- Host: GitHub
- URL: https://github.com/eleutherai/features-across-time
- Owner: EleutherAI
- License: mit
- Created: 2023-12-20T21:26:04.000Z (over 2 years ago)
- Default Branch: main
- Last Pushed: 2024-04-12T05:25:36.000Z (about 2 years ago)
- Last Synced: 2024-04-12T13:18:58.632Z (about 2 years ago)
- Language: Python
- Size: 7.93 MB
- Stars: 24
- Watchers: 3
- Forks: 1
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# features-across-time
Understanding how features learned by neural networks evolve throughout training
NOTE: Apologies for the sloppy code in many places. We plan to clean it up in the next few weeks.
### Reproduce Pythia Suite Experiments
Run the following Python scripts from the command line to collect and plot data over training steps:
```
python -m scripts.preprocess.build_ngram_data --n
python -m scripts.inference.ngrams --n
```
To plot existing data, use:
```
python -m scripts.plot.plot_single_rows
```