⚠️ Repository archived

This repository is archived and no longer maintained.
Please move to the active repository: https://github.com/yophis/partial-yarn

If you landed here from an older link, update your remotes:

git remote set-url origin https://github.com/yophis/partial-yarn.git

---

Partial YaRN: Audio-Only Context Extension for Large Audio-Language Models

[Paper] | [Dataset]

This repository contains the PyTorch code of Partial YaRN and Virtual Longform Audio Training (VLAT) from the paper:
Extending Audio Context for Long-Form Understanding in Large Audio-Language Models. 2025.

Our work focuses on extending the audio context window of Large Audio-Language Models (LALMs) without degrading the sophisticated language capabilities of the base Large Language Model (LLM). The core idea is to apply context extension methods, such as YaRN, in a targeted manner. Instead of altering the positional encodings of the entire input sequence, Partial YaRN modifies only the audio tokens. This preserves the original positional information for text tokens, thereby protecting the LLM's pretrained language understanding.