Add descriptive dataset statistics plots

[yananlong]

Summary

• add descriptive dataset statistics helpers for model coverage by benchmark and inference-engine/runtime spread
• add PDF plots for model-per-dataset coverage and inference-engine spread
• cover the new descriptive report fields with focused tests
• keep the plotting script focused on PDF output only

Validation

• .venv/bin/python -m pytest tests/test_dataset_statistics.py
• .venv/bin/python scripts/plot_dataset_statistics.py --help

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[tommasocerruti]

I have just one question about how the score summaries are grouped, then this looks good :)

[tommasocerruti]

Should these grouped score summaries include a metric identity field too (e.g., metric_id)? The code’s benchmark field comes from source_data.dataset_name in the datastore (as you defined in line 60 of this file), and some evaluations in the datastore report multiple metrics under the same dataset_name + evaluation_name pair. For example, in this arc-agi file, ARC Prize evaluations leaderboard JSON + v1_Semi_Private is used for both score (accuracy) and cost_per_task (cost). I believe grouping only by benchmark + evaluation_name would average those different quantities together, do you agree?

[yananlong]

This sounds reasonable. Let me investigate further.

[tommasocerruti]

@yananlong are you still working on this, or can I start reviewing it?

[yananlong]

Still working. I am making different plots now. More on this later today.

[tommasocerruti]

Great, feel free to ping me when you are done.

[nelaturuharsha]

Is this still WIP @yananlong ?