sail-sg/Attention-Sink

[ICLR 2025] When Attention Sink Emerges in Language Models: An Empirical View (Spotlight)

35
/ 100
Emerging

This project helps researchers and developers understand how 'attention sink' behavior emerges in large language models (LLMs) during pre-training. By providing tools to analyze factors like optimization, data, and architecture, it allows users to examine attention patterns in open-source LLMs or their own pre-trained models. The output helps machine learning researchers diagnose and interpret LLM training dynamics.

159 stars. No commits in the last 6 months.

Use this if you are a machine learning researcher or engineer actively pre-training or analyzing the internal workings of large language models and want to investigate the phenomenon of attention sink.

Not ideal if you are an end-user simply looking to apply or fine-tune existing large language models without delving into their pre-training mechanics.

large-language-models LLM-pretraining machine-learning-research attention-mechanisms model-analysis
Stale 6m No Package No Dependents
Maintenance 2 / 25
Adoption 10 / 25
Maturity 16 / 25
Community 7 / 25

How are scores calculated?

Stars

159

Forks

5

Language

Python

License

MIT

Last pushed

Jul 08, 2025

Commits (30d)

0

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/transformers/sail-sg/Attention-Sink"

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