shawnmkim/compete-collab-taskbank

Compete-then-Collaborate Task Bank

Companion dataset for the paper "Compete then Collaborate: Frontier AI Teachers Build a Verifiable Curriculum to Improve a Coding Student Beyond Imitation" (Kim, 2026).

This dataset contains the execution-verified task bank used to (a) rank four frontier LLM teachers (Claude, Codex-GPT, Grok, Gemini) and (b) train the coding student model Qwen2.5-Coder-7B via imitation (SFT) and verifiable-reward RL (RLVR/GRPO).

• 📄 Paper: https://github.com/shawnkim678/compete-then-collaborate
• 🤖 Companion RLVR model: shawnmkim/qwen2.5-coder-7b-rlvr-compete-collab
• 📦 Code / verifier / harness: https://github.com/shawnkim678/compete-then-collaborate

Splits & sizes

Config	Split	Rows	Size	Source	Held-out from training?
mbpp	train	~350	106 KB	google-research-datasets/mbpp (teaching split)	training
mbpp	heldout	150	76 KB	MBPP test split	held-out for evaluation
contests	train	~150	383 KB	deepmind/code_contests d6–9	training / RLVR reward env.
contests	heldout	68	254 KB	disjoint subset from above	held-out for evaluation

Leakage verified: the held-out splits are disjoint from the training splits at the task_id level.

Data schema

mbpp split — function tasks with unit-test asserts

{
  "task_id":     "mbpp_11",
  "category":    "function",
  "instruction": "Write a function that ...",
  "input":       "<function signature or example>",
  "tests":       ["assert f(1) == 2", "assert f(3) == 5", ...]
}

contests split — competition problems with stdin/stdout tests

{
  "task_id":     "cc_1234",
  "category":    "competition",
  "difficulty":  7,
  "instruction": "<problem statement>",
  "tests_io":    [{"stdin": "3\n1 2 3\n", "stdout": "6\n"}, ...]
}

Load (🤗 datasets)

from datasets import load_dataset

mbpp_train   = load_dataset("shawnmkim/compete-collab-taskbank", "mbpp",     split="train")
mbpp_heldout = load_dataset("shawnmkim/compete-collab-taskbank", "mbpp",     split="heldout")
cc_train     = load_dataset("shawnmkim/compete-collab-taskbank", "contests", split="train")
cc_heldout   = load_dataset("shawnmkim/compete-collab-taskbank", "contests", split="heldout")

How it was built

1. MBPP source (Google, Apache-2.0). Function-signature disambiguated; externally-dependent tasks excluded to keep function-only comparison valid. Teaching / held-out split separated with a fixed seed.
2. code_contests source (DeepMind, CC-BY-4.0). Restricted to difficulty 6–9 ("hard"), with stdin/stdout tests preserved verbatim.
3. Intersection control (paper §3.2): the student training set is further restricted to problems all teachers solved, giving identical problems and equal counts across teachers so only solution / explanation style differs.

The full build scripts are in the GitHub repo (build_taskbank_mbpp.py, build_contests_bank.py).

Reward for RLVR

For the RLVR (GRPO) reward computation, we execute the student's generated code in an isolated subprocess (rlimits + timeout) against the hidden tests / tests_io fields. Reward = fraction of tests passed + small format bonus. See paper §3.3 and §4.

License & attribution

This dataset is a derivative curation of two upstream sources; the strictest compatible terms apply:

• MBPP-derived rows: Apache-2.0 (Google Research)
• code_contests-derived rows: CC-BY-4.0 (Google DeepMind)

Please cite both the paper (below) and the upstream sources when using.

Citation

@misc{kim2026compete,
  title  = {Compete then Collaborate: Frontier AI Teachers Build a Verifiable Curriculum to Improve a Coding Student Beyond Imitation},
  author = {Kim, Miseong (Shawn)},
  year   = {2026},
  note   = {Preprint, Genesis Cortex AI Inc.},
  url    = {https://github.com/shawnkim678/compete-then-collaborate}
}

@article{austin2021mbpp,
  title   = {Program Synthesis with Large Language Models},
  author  = {Austin, Jacob and Odena, Augustus and Nye, Maxwell and others},
  journal = {arXiv preprint arXiv:2108.07732},
  year    = {2021}
}

@article{li2022codecontests,
  title   = {Competition-level code generation with AlphaCode},
  author  = {Li, Yujia and Choi, David and Chung, Junyoung and others},
  journal = {Science},
  year    = {2022}
}

Ethics / ToS note

No teacher raw outputs are redistributed. This dataset consists exclusively of task specifications and hidden tests derived from publicly licensed sources (MBPP · code_contests). Users regenerate teacher solutions with their own API keys under their own provider terms.