thughost's activity

• All
• Models
• Datasets
• Spaces
• Papers
• Collections
• Community
• Posts
• Upvotes
• Likes
• Articles

posted an update over 1 year ago

view post

Post

756

We've open-sourced the code and models for Self-Play Preference Optimization (SPPO)! 🚀🚀🚀
🤗paper: Self-Play Preference Optimization for Language Model Alignment (2405.00675)
⭐ code: https://github.com/uclaml/SPPO
🤗models: UCLA-AGI/sppo-6635fdd844f2b2e4a94d0b9a

🔥

3

3

+

Reply

posted an update almost 2 years ago

view post

Post

Check out the demo of SPIN-Diffusion made by @angelahzyuan at: UCLA-AGI/SPIN-Diffusion-demo-v1

• 
• 45 replies

·

❤️

5

5

+

Reply

reacted to akhaliq's post with 👍 almost 2 years ago

view post

Post

LongRoPE

Extending LLM Context Window Beyond 2 Million Tokens

LongRoPE: Extending LLM Context Window Beyond 2 Million Tokens (2402.13753)

Large context window is a desirable feature in large language models (LLMs). However, due to high fine-tuning costs, scarcity of long texts, and catastrophic values introduced by new token positions, current extended context windows are limited to around 128k tokens. This paper introduces LongRoPE that, for the first time, extends the context window of pre-trained LLMs to an impressive 2048k tokens, with up to only 1k fine-tuning steps at within 256k training lengths, while maintaining performance at the original short context window. This is achieved by three key innovations: (i) we identify and exploit two forms of non-uniformities in positional interpolation through an efficient search, providing a better initialization for fine-tuning and enabling an 8x extension in non-fine-tuning scenarios; (ii) we introduce a progressive extension strategy that first fine-tunes a 256k length LLM and then conducts a second positional interpolation on the fine-tuned extended LLM to achieve a 2048k context window; (iii) we readjust LongRoPE on 8k length to recover the short context window performance. Extensive experiments on LLaMA2 and Mistral across various tasks demonstrate the effectiveness of our method. Models extended via LongRoPE retain the original architecture with minor modifications to the positional embedding, and can reuse most pre-existing optimizations.

• 

• 
• 
• 3 replies

·

👍

13

13

🤯

8

8

+

Reply