[pull] main from huggingface:main

[pull]

See Commits and Changes for more details.

---

Created by pull[bot] (v2.0.0-alpha.1)

Can you help keep this open source service alive? 💖 Please sponsor : )

Summary by Sourcery

This pull request focuses on improving the reproducibility of Deepseek's evaluation results and updating dependencies. It introduces new prompt templates for MATH and GPQA tasks, and aligns the evaluation settings with the DeepSeek-R1 paper by using sampling with temperature 0.6 and top_p 0.95. The README is updated with detailed instructions and commands to reproduce the results, along with tables comparing the reproduced results with the reported results from Deepseek. Also, the dependencies in setup.py are updated to specific versions.

Bug Fixes:

• Fixes discrepancies between reported and reproduced results by using sampling with temperature 0.6 and top_p 0.95, aligning with the DeepSeek-R1 paper's evaluation settings.
• Updates the documentation to reflect the correct parameter name max_new_tokens instead of generation_size.

Enhancements:

• Improves the accuracy of reproducing Deepseek's evaluation results on AIME 2024, MATH-500, and GPQA Diamond benchmarks.
• Adds support for LiveCodeBench evaluations.

Documentation:

• Updates the README with detailed instructions and commands to reproduce Deepseek's evaluation results on AIME 2024, MATH-500, and GPQA Diamond benchmarks.
• Adds tables comparing the reproduced results with the reported results from Deepseek for AIME 2024, MATH-500, and GPQA Diamond benchmarks.

Tests:

• Updates the prompt templates for MATH and GPQA tasks to improve evaluation accuracy.

Chores:

• Updates the dependencies in setup.py to specific versions, including accelerate, transformers, and trl.

[gitstream-cm]

🚨 gitStream Monthly Automation Limit Reached 🚨

Your organization has exceeded the number of pull requests allowed for automation with gitStream.
Monthly PRs automated: 1998/250

To continue automating your PR workflows and unlock additional features, please contact LinearB.

[sourcery-ai]

Reviewer's Guide by Sourcery

This pull request updates the evaluation setup to align with DeepSeek's reported results, introduces standardized prompt templates for math and GPQA tasks, and updates dependencies to specific versions.

Sequence diagram for the evaluation process

sequenceDiagram
    participant User
    participant lighteval
    participant vllm
    participant Model

    User->>lighteval: Runs evaluation with specified MODEL_ARGS (temperature=0.6, top_p=0.95)
    lighteval->>vllm: Sends evaluation request with prompt
    vllm->>Model: Generates response based on prompt
    Model-->>vllm: Returns generated response
    vllm-->>lighteval: Returns response
    lighteval->>lighteval: Evaluates response
    lighteval-->>User: Returns evaluation results

Loading

Updated class diagram for LightevalTaskConfig

classDiagram
    class LightevalTaskConfig {
        +name: str
        +suite: list[str]
        +prompt_function: function
        +hf_repo: str
        +hf_subset: str
        +hf_avail_splits: list[str]
        +metric_function: function
    }

    note for LightevalTaskConfig "The prompt_function attribute now uses standardized prompt templates for math and GPQA tasks."

Loading

File-Level Changes

Change	Details	Files
Updates the evaluation commands in README.md to use a temperature of 0.6 and a top-p value of 0.95 for sampling, aligning with the DeepSeek-R1 paper's settings.	Updates the MODEL_ARGS in the evaluation commands to include temperature:0.6,top_p:0.95. Updates the reproduction notes to reflect the sampling strategy.	README.md
Introduces prompt templates for math and GPQA tasks in evaluate.py to standardize the input format for models.	Adds MATH_QUERY_TEMPLATE for math problems, enforcing a specific output format. Adds GPQA_QUERY_TEMPLATE for multiple-choice questions, enforcing a specific output format. Updates math_prompt_fn to use the new MATH_QUERY_TEMPLATE. Updates gpqa_prompt_fn to use the new GPQA_QUERY_TEMPLATE.	src/open_r1/evaluate.py
Updates the dependencies in setup.py to specific versions.	Updates accelerate to version 1.4.0. Updates transformers to version 4.49.0. Updates trl to a specific commit hash.	setup.py

---

Tips and commands

Interacting with Sourcery

• Trigger a new review: Comment @sourcery-ai review on the pull request.
• Continue discussions: Reply directly to Sourcery's review comments.
• Generate a GitHub issue from a review comment: Ask Sourcery to create an
  issue from a review comment by replying to it. You can also reply to a
  review comment with @sourcery-ai issue to create an issue from it.
• Generate a pull request title: Write @sourcery-ai anywhere in the pull
  request title to generate a title at any time. You can also comment
  @sourcery-ai title on the pull request to (re-)generate the title at any time.
• Generate a pull request summary: Write @sourcery-ai summary anywhere in
  the pull request body to generate a PR summary at any time exactly where you
  want it. You can also comment @sourcery-ai summary on the pull request to
  (re-)generate the summary at any time.
• Generate reviewer's guide: Comment @sourcery-ai guide on the pull
  request to (re-)generate the reviewer's guide at any time.
• Resolve all Sourcery comments: Comment @sourcery-ai resolve on the
  pull request to resolve all Sourcery comments. Useful if you've already
  addressed all the comments and don't want to see them anymore.
• Dismiss all Sourcery reviews: Comment @sourcery-ai dismiss on the pull
  request to dismiss all existing Sourcery reviews. Especially useful if you
  want to start fresh with a new review - don't forget to comment
  @sourcery-ai review to trigger a new review!
• Generate a plan of action for an issue: Comment @sourcery-ai plan on
  an issue to generate a plan of action for it.

Customizing Your Experience

Access your dashboard to:

• Enable or disable review features such as the Sourcery-generated pull request
  summary, the reviewer's guide, and others.
• Change the review language.
• Add, remove or edit custom review instructions.
• Adjust other review settings.

Getting Help

• Contact our support team for questions or feedback.
• Visit our documentation for detailed guides and information.
• Keep in touch with the Sourcery team by following us on X/Twitter, LinkedIn or GitHub.