Datasets:

eltociear
/

csapp-tasks-v1

license: mit
task_categories:
  - text-generation
tags:
  - computer-systems
  - csapp
  - bit-manipulation
  - floating-point
  - cache
  - code
  - rl-environment
size_categories:
  - n<1K

csapp-tasks-v1

40 machine-level computer-systems tasks for the csapp-env RL environment — two's complement, integer overflow, IEEE-754, bit manipulation, struct layout and caches, on the topics of Computer Systems: A Programmer's Perspective (Bryant & O'Hallaron).

field	meaning
`task_id`	`cs-000` … `cs-039`
`category`	integers / floats / bits / memory / cache
`prompt`	the question and the exact shape of the answer
`api_description`	the shared preamble (available imports, how to emulate C's fixed widths)
`expected_output`	JSON `{"rows": [...]}`, computed by executing a reference solution

Categories: integers 12, floats 10, bits 6, memory 6, cache 6. No dependencies — every task is answered with struct and math.

Original problems, not the book's exercises

Nothing here is copied from CS:APP. Every task is written fresh on the topics the book teaches.

That is a copyright decision and the stronger evaluation choice: the book's exercises and their worked solutions are all over the public web, so a model can recall them instead of computing. Original problems with executed answer keys cannot be recalled.

The answer keys are executed, never typed

Every expected output is produced by running a reference implementation, so a key cannot drift from the instruction beside it. That rule caught four defects in this dataset's own first draft:

a task that said "show that float32 addition is not associative" where both groupings came out 0.0 — the values chosen did not demonstrate it;
a row-major vs column-major cache comparison where both orders tied at 8, because the cache was large enough to hold the array;
a hit/miss trace that was five misses and no hits, so answering 0 five times scored full marks;
struct offsets that were four typed constants instead of a computed layout.

All four are now asserted in the builder, so the dataset fails to build if any of them recurs.

Grading

Exact — no tolerance anywhere. Every answer is an integer, a bit pattern, a boolean or a float read back from its own bit pattern, so there is nothing for a tolerance to absorb, and on a question about exact bit patterns a tolerance would only let wrong answers through. 1 does not pass for True, and "0x2a" does not pass for 42.

Verify with python environments/csapp_env/build_tasks.py --verify (40/40). Source: https://github.com/eltociear/my-molt-agent/tree/main/environments/csapp_env