Spaces:

open-llm-leaderboard

/

open_llm_leaderboard

like 11.7k

Running on CPU Upgrade

Changes in app.py

• Search Model Function: Updated to search within AutoEvalColumn.model.name instead of AutoEvalColumn.dummy.name.
• Select Columns Function: Removed references to AutoEvalColumn.dummy.name and adjusted to only include necessary columns.
• Dataframe Component in Gradio Interface: Updated to ensure that the table no longer includes AutoEvalColumn.dummy.name.

Changes in css_html_js.py

• CSS Rules: Removed CSS rules that hid the last column (which was the dummy column).

Changes in utils.py

• Column Definitions: The main thing – refactored the way columns are defined and initialized. Removed the initialization that directly added the dummy column.
• Dataclass Initialization: Updated to create the AutoEvalColumn dataclass without the dummy column.

Changes in filter_models.py

• Flag Models Function: Modified to use AutoEvalColumn.model.name for flagging logic instead of using a separate "model_name_for_query" which was tied to the dummy column.
• Remove Forbidden Models Function: Updated to check against AutoEvalColumn.model.name instead of a non-existent "model_name_for_query".

Changes in read_evals.py

• EvalResult to_dict Method: Removed the line that added the dummy column's data to the dictionary. Now it only uses relevant and existing columns.

Changes in collections.py

• Update Collections Function: Updated the sorting and selection logic to use AutoEvalColumn.model.name instead of AutoEvalColumn.dummy.name.

Functionality

I have tested that the search and buttons are working correctly. I'll need your review @clefourrier :3

Hello!

Looking good!

Some specific comments to address before merging:

• src/display/utils.py - Doesn't your new system reorder the columns in the table?
• src/leaderboard/filter_models.py. You remove

# Merges and moes are flagged automatically
        if model_data[AutoEvalColumn.flagged.name]:
            flag_key = "merged"

which means that some models (relying on the tags and dynamic data for flagging) will no longer be flagged I think.

• appends in src/display/utils.py are my long-standing problem, each time I'm trying to make them more readable. You're right, this idea reorder the column in the table, let revert this change.
• my bad with src/leaderboard/filter_models.py 😱

Besides, I see, if we don't use this dummy column so we don't have a simple model name, and AutoEvalColumn.model.name contains specs, for example, this one:
model check: <a target="_blank" href="https://huggingface.co/ceadar-ie/FinanceConnect-13B" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">ceadar-ie/FinanceConnect-13B</a> <a target="_blank" href="https://huggingface.co/datasets/open-llm-leaderboard/details_ceadar-ie__FinanceConnect-13B" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">📑</a>

And when I removed the dummy column and tried to filter by AutoEvalColumn.model.name, the flags filter breaks. It's manageable, let me think.

So, the dummy column appeared to be a useful one, but let's rename it to "fullname", it's a little bit more clear. Three thoughts:

1. I should've called this PR "There and Back Again", but we found out the model name is not a simple model name.
2. We need to refactor read_evals.py because there're lots data transformation right now, we need to simplify it.
3. We need to do smth with model name fields, it's currently a mess.

PR is now merged/closed. The ephemeral Space has been deleted.
(This is an automated message.)