Skip to content

Can not reproduce the results on LongBench in the paper. #1

@nietzhuang

Description

@nietzhuang

Thank you for your great work on KV cache merging, D2O's dynamic token merging method inspires me a lot.

Currently I'm running your source code.
As stated in the paper, I choose N:M=3:1, alpha=0.3 and beta=0.7 under 20% (rho=0.2) KV cache compression ratio.
I'm sure that the Python dependancies and environment are met with your source code, and I'm using nvidia RTX PRO6000 GPU, which can run LongBench properly.

However, I found the results are very different from the paper.
The results from source code may not show D2O's strength.
If there are something I misunderstood or may omit, could you help to solve this issue?
Thank you very much.

Results from paper show,
| H2O | D2O
NarrativeQA | 13.27 | 14.43
Qasper | 11.05 | 12.66
MF-en | 17.72 | 19.93
HotpotQA | 10.38 | 11.92
2WikiMQA | 11.23 | 12.79
Musique | 6.38 | 9.88
GovReport | 21.29 | 24.36
QMSum | 21.33 | 23.42
MultiNews | 3.38 | 3.95
TREC | 66.63 | 69.72
TriviaQA | 89.19 | 90.99
SAMSum | 41.12 | 42.36
Pcount | 5.52 | 6.61
Pre | 11.11 | 14.67
Lcc | 71.86 | 72.43
RB-P | 58.29 | 60

Results from source code show,
| H2O | D2O
NarrativeQA | 12.43 | 12.64
Qasper | 12.55 | 11.92
MF-en | 19.95 | 19.87
HotpotQA | 10.92 | 10.72
2WikiMQA | 12.2 | 11.95
Musique | 6.65 | 6.75
GovReport | 22.97 | 21.13
QMSum | 23.44 | 23.13
MultiNews | 3.56 | 1.93
TREC | 69 | 69.67
TriviaQA | 90.57 | 90.63
SAMSum | 41.96 | 42.05
Pcount | 5.18 | 5.9
Pre | 11.58 | 13.86
Lcc | 69.26 | 71.57
RB-P | 55.67 | 58.43

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions