Yongkang Liu

[Yongkang Liu @liuygkg]

@HaritzPuerto I have also encountered similar reviewers who are very irresponsible. I hope ACL will be able to restrain and penalize bad reviewers in the future. @aclmeeting

[Yongkang Liu @liuygkg]

@NielKlug @ECMLPKDD @CisLmu @HinrichSchuetze Congrats Ercong!

[Ercong Nie @NielKlug]

🥳Delighted to announce that our work "A Unified Data Augmentation Framework for Low-Resource Multi-Domain Dialogue Generation" has been accepted to @ECMLPKDD 2024! @CisLmu
Sincere thanks to my collaborators and supervisors! @liuygkg @HinrichSchuetze, and all others not on X.

[Yongkang Liu @liuygkg]

HiFT is open-source! It's a model-independent optimizer-independent full-parameter fine-tuning method that supports full-parameter fine-tuning of 7B models on 24G memory devices. It's based on huggingface and can be used by "pip install hift"
github link: https://t.co/CDmY1DbWBN

[Yongkang Liu @liuygkg]

Happy to share our new paper "HiFT: A Hierarchical Full Parameter Fine-Tuning Strategy".
📄(https://t.co/k4AvPiWznl).

👥 joint work with Yiqun Zhang, Qian Li, Shi Feng, Daling Wang, Yifei Zhang and @HinrichSchuetze.

[Yongkang Liu @liuygkg]

@llamafactory_ai @Clear_Fing We are very happy to see such interesting work. This work is very similar to our existing work [1] and uses almost the same optimization process. But there seems to be no comparison or explanation in the paper.

HiFT: A Hierarchical Full Parameter Fine-Tuning Strategy

[Yongkang Liu @liuygkg]

For the 13B model, the peak GPU memory requirement is 31GB.
The up-to-date results are based on mixed-precision for adaptation hierarchy strategies.

[Yongkang Liu @liuygkg]

Happy to share our new paper "HiFT: A Hierarchical Full Parameter Fine-Tuning Strategy".
📄(https://t.co/k4AvPiWznl).

👥 joint work with Yiqun Zhang, Qian Li, Shi Feng, Daling Wang, Yifei Zhang and @HinrichSchuetze.

[Yongkang Liu @liuygkg]

The latest experimental results show that HiTF enables full-parameter fine-tuning of a 7B model on devices with 24G memory with a single precision using the AdamW optimizer. When the batch is 1 and the sentence length is 512, the peak memory requirement is about 16.8G.

[Yongkang Liu @liuygkg]

HiFT is an end-to-end optimizer-independent hierarchical fine-tuning strategy that greatly reduces the memory requirements of full-parameter fine-tuning language models.

[Yongkang Liu @liuygkg]

Note that results in the current paper are based on mixed-precision unadapted hierarchical strategy.

[Yongkang Liu @liuygkg]

@LrecColing Why are appendices not allowed during submission? It’s incomprehensible! I am afraid that work such as pre-training models cannot be submitted to LREC-COLING.

[Yongkang Liu @liuygkg]

@imani_ayyoob @CisLmu @lpq29743 Congrats!

[Yongkang Liu @liuygkg]

@lpq29743 Congrats Peiqin!!!

[Yongkang Liu @liuygkg]

🥳Thrilled to announce that "PVGRU: Generating Diverse and Relevant Dialogue Responses via Pseudo-Variational Mechanism" has been accepted to #ACL2023!
👥joint work with: Shi Feng, Daling Wang, Yifei Zhang
and @HinrichSchuetze.
paper:https://t.co/cem1TLzjUG

[Yongkang Liu @liuygkg]

@annargrs @rikvannoord Do you think it is reasonable to not allow 0.5 for soundness?

[Yongkang Liu @liuygkg]

@annargrs @yanzuxu @aclmeeting Great！

[Yongkang Liu @liuygkg]

@rikvannoord @annargrs @aclmeeting I also think this is an important issue, soundness should allow 0.5 to appear, the gap between papers is not that big.

[Yongkang Liu @liuygkg]

@annargrs @aclmeeting Yes, the bug has been fixed.