Changes related to running benchmark experiments for the paper: Support Qwen3.5 and thinking models, Skywork, truncation tracking, benchmark changes etc#32
Open
ErlisLushtaku wants to merge 24 commits intomainfrom
Conversation
- fix dependencies - add structured output to prevent judge from not respecting the prompt
kargibora
reviewed
Apr 7, 2026
| [project.optional-dependencies] | ||
| vllm = ["vllm==0.10.2", "transformers>=4.55.2,<5.0.0"] | ||
| # vLLM on PyPI pins transformers<5; optional extra matches that so `uv lock` can resolve. | ||
| vllm = ["vllm>=0.17.0,<1.0.0", "transformers>=4.56.0,<5.0.0"] |
Collaborator
There was a problem hiding this comment.
vllm>=0.17.0,<1.0.0 is a very wide range. A few concerns:
- Was this tested with a prebuilt wheel or built from source? Building vLLM from source on cluster nodes often fails due to CUDA kernel compilation issues.
- Is the
StructuredOutputsParamsimport path (vllm.sampling_params) stable across this entire range? It may have been introduced in 0.17 and could move. For exampleStructuredOutputParamswas a bit different whenvllm==0.11.0. Thus I think it makes more sense to create more stable versioning
Collaborator
Author
There was a problem hiding this comment.
Good point. I tightened the range. 0.18.1 was working. I think the StructuredOutputParams is stable accross the new range.
Collaborator
Author
There was a problem hiding this comment.
Changed it to v0.19+ so that we can use the thinking token limit parameter, and also they have some fixes for Qwen3.5
ErlisLushtaku
force-pushed
the
erlislushtaku/fix/support-qwen-3.5
branch
from
ab3db1b to
ef1c92c
Compare
- Switch from choice-based structured outputs to JSON schema constraint - Tighten vllm version range from >=0.17.0,<1.0.0 to >=0.17.0,<0.19.0
…ench baseline from huggingface and update huggingface repo
ErlisLushtaku
changed the title
…gex stripping since the structured output wasn't working for isolating thinking tokens anyway
ErlisLushtaku
changed the title
ErlisLushtaku
commented
Apr 17, 2026
…so that we have more customizability - Introduced `truncate_judge_input_chars` and `max_judge_model_len` to `BaseCliArgs` for better control over judge-side input limits.
- Refactor baseline assignment for Arena-Hard datasets to support different baselines based on category same as original benchmark.
ErlisLushtaku
changed the title
…ted token count for max_model_len
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Uh oh!
There was an error while loading. Please reload this page.