Skip to content

Conversation

XiaotongJiang
Copy link
Collaborator

@XiaotongJiang XiaotongJiang commented Dec 30, 2024

Issue #2504

  • Compared SGL and vLLM's math eval score to address concerns on SGL being constant low on this eval (res linked below)
  • Add math eval to CI

Setup:

  • Prompt: simple eval from openai
  • Grading Model: gemini-1.5-flash
  • Temperature 0.0
inference engine Model # Examples score model official score
vllm llama 3.2 1b 500 0.248 0.306
sgl llama 3.2 1b 500 0.248 0.306
vllm llama 3.2 3b 500 0.476 0.48
sgl llama 3.2 3b 500 0.484 0.48
vllm llama 3.1 8b 500 0.50 0.519
sgl llama 3.1 8b 500 0.496 0.519
sgl llama 3.1 8b 65 0.492 0.519

@XiaotongJiang XiaotongJiang changed the title Add math eval to CI [Issue 2504] Add math eval to CI Dec 30, 2024
@zhyncs zhyncs changed the title [Issue 2504] Add math eval to CI [feat] Add math eval to CI Dec 30, 2024
@zhyncs zhyncs merged commit a11f8d5 into sgl-project:main Dec 30, 2024
@zhyncs
Copy link
Member

zhyncs commented Dec 30, 2024

Thanks!!

@zhaochenyang20
Copy link
Collaborator

@XiaotongJiang @zhyncs @merrymercy @shuaills I am talking with XIaotong and will take a look on this. Also, thanks so much. this having been suspended for so long 😂

merrymercy added a commit that referenced this pull request Dec 30, 2024
@XiaotongJiang
Copy link
Collaborator Author

XiaotongJiang commented Dec 30, 2024

sorry for the disruption

AssertionError: 0.505 not greater than or equal to 0.509

need to loose this assertion a bit

XiaotongJiang added a commit to XiaotongJiang/sglang that referenced this pull request Jan 3, 2025
timethink pushed a commit to timethink/sglang that referenced this pull request Mar 9, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants