Hugging Face's logo

Datasets:
osunlp
/
ScienceAgentBench

Modalities:
Text
Languages:
English
Size:
< 1K
ArXiv:
License:
Dataset card Data Studio Files
xet
 Community
3

Task 77 (verified split): output file name typo in the instruction (“interploated” vs “interpolated”)

#3
by james-sullivan - opened
Discussion
james-sullivan

Hi, and thanks for releasing the verified split.

In the verified split, instance_id 77 (gold program WaterQuality.py, eval script WaterQuality_eval.py) has a file-name mismatch between the instruction and the files that grade it:

  • task_inst asks the program to save the map as "pred_results/interploated_water_quality.png" (the "o" and "l" in "interpolated" are swapped).
  • output_fname is pred_results/interpolated_water_quality.png.
  • benchmark/eval_programs/WaterQuality_eval.py (line 4) reads pred_results/interpolated_water_quality.png.
  • benchmark/gold_programs/WaterQuality.py (line 43) saves pred_results/interpolated_water_quality.png.

So a program that follows the instruction exactly writes a file the eval script never reads, and it is scored as not saving its output. We saw this happen: an agent that followed the instruction literally scored 0 on this task. The same typo is also in the older validation config (ScienceAgentBench.csv).

Suggested fix: change "interploated" to "interpolated" in task 77's task_inst (in both configs, if you keep the old one).

We checked all 102 tasks for this kind of instruction / output-file mismatch and found only this one.

Written by Claude

Sign up or log in to comment