liutianlin0121/decoding-time-realignment

Implementation of "Decoding-time Realignment of Language Models", ICML 2024.

25
/ 100
Experimental

This tool helps AI engineers and researchers fine-tune how aligned a language model is with specific user preferences or applications, without needing to retrain the model. You provide an existing RLHF-aligned language model, and it allows you to adjust its alignment strength during the decoding (generation) process. This is for AI practitioners working on deploying or optimizing large language models.

No commits in the last 6 months.

Use this if you need to quickly adjust the alignment behavior of an already trained, RLHF-aligned language model for different use cases or to find optimal regularization strengths for future retraining, without the time and cost of full retraining.

Not ideal if you are looking to train a new language model from scratch or if your model is not already aligned using Reinforcement Learning from Human Feedback (RLHF).

AI-model-alignment Large-Language-Models NLP-fine-tuning ML-experimentation AI-safety-tuning
No License Stale 6m No Package No Dependents
Maintenance 0 / 25
Adoption 6 / 25
Maturity 8 / 25
Community 11 / 25

How are scores calculated?

Stars

21

Forks

3

Language

Jupyter Notebook

License

Last pushed

Jun 17, 2024

Commits (30d)

0

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/nlp/liutianlin0121/decoding-time-realignment"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.