Updates experiment script

[finbarrtimbers]

No description provided.

[gemini-code-assist]

Code Review

This pull request updates the qwen3_4b_dapo_math.sh script to allow the Beaker image to be set via a positional argument, modifies default cluster and priority settings, and updates evaluation datasets to the allenai namespace. A significant issue was identified in the logic for parsing the Beaker image argument, as it could incorrectly consume command-line flags intended for the training script.

[finbarrtimbers]

Switched to passing BEAKER_IMAGE via env var from build_image_and_launch.sh per @mnoukhov's suggestion. The qwen script now just reads ${BEAKER_IMAGE:-nathanl/open_instruct_auto} with no positional consumption, so flags like --learning_rate pass through cleanly.

[mnoukhov]

lgtm, let's do the BEAKER_IMAGE change in a future PR