README.md

metadata

license: mit
task_categories:
  - text-generation
  - text-classification
language:
  - en
pretty_name: IntentSpec Benchmark
size_categories:
  - n<1K
configs:
  - config_name: spec_pairs
    default: true
    data_files:
      - split: train
        path: spec_pairs.jsonl
  - config_name: canonicals
    data_files:
      - split: train
        path: canonicals.jsonl

IntentSpec Benchmark — Data Supplement

This archive contains the benchmark data used to compute Intent Violation Rate (IVR) in the paper: 49 tasks, each derived from a HumanEval problem and extended with an ambiguous/gold prompt pair and a decomposed set of executable constraints.

Files

spec_pairs.jsonl

The benchmark itself — one JSON object per line, one line per task. This is the file consumed directly by the evaluation pipeline (run_experiment1.py, validate_benchmark.py in the software supplement).

Fields:

Field	Type	Description
task_id	string	HumanEval task identifier, e.g. "HumanEval/26". Matches the task_id in canonicals.jsonl.
spec_type	string	Category of the task. Currently always "algorithm".
ambiguous_prompt	string	The under-specified prompt a developer might plausibly write — states only some of the task's real constraints.
gold_prompt	string	The fully clarified version of the prompt, stating every constraint explicitly.
constraints	list of objects	All atomic, executable constraints implied by the gold prompt (see below).
stated_test_ids	list of strings	IDs (from constraints) that are mentioned, even implicitly, in ambiguous_prompt.
hidden_test_ids	list of strings	IDs (from constraints) that appear only in gold_prompt, not in ambiguous_prompt. A solution that passes all stated_test_ids but fails one or more hidden_test_ids counts as an intent violation — this is what IVR measures.

Each entry in constraints is an object:

Field	Type	Description
id	string	Local constraint identifier (e.g. "C1"), referenced by stated_test_ids/hidden_test_ids.
description	string	Human-readable statement of the constraint.
test_code	string	Executable Python snippet that calls solution(...) and asserts the constraint holds. Run by concatenating a candidate solution with this snippet and executing it; a zero exit code means the constraint passed.

Example (abbreviated):

{
  "task_id": "HumanEval/26",
  "spec_type": "algorithm",
  "ambiguous_prompt": "Write a function named `solution` that takes a list of integers and removes any elements that appear more than once.",
  "gold_prompt": "Write a function named `solution` that takes a list of integers, removes all elements that occur more than once, preserves the original order of remaining elements, and returns a new list without mutating the input.",
  "constraints": [
    {"id": "C1", "description": "Elements appearing more than once are removed", "test_code": "result = solution([1, 2, 3, 2, 4])\nassert 2 not in result"},
    {"id": "C2", "description": "Original order of remaining elements is preserved", "test_code": "..."},
    {"id": "C3", "description": "Input list is not mutated", "test_code": "..."},
    {"id": "C4", "description": "Empty list returns empty list", "test_code": "..."}
  ],
  "stated_test_ids": ["C1"],
  "hidden_test_ids": ["C2", "C3", "C4"]
}

canonicals.jsonl

The original HumanEval reference (gold) solutions that spec_pairs.jsonl was built from. Used for benchmark validation — confirming that a correct, known-good solution passes every constraint (stated and hidden) for its task, i.e. that IVR == 0 for canonical solutions.

Fields:

Field	Type	Description
task_id	string	HumanEval task identifier, matches spec_pairs.jsonl.
entry_point	string	Name of the function under test in the original HumanEval problem.
prompt	string	Original HumanEval function signature + docstring.
canonical_solution	string	Original HumanEval reference implementation body.
full_source	string	prompt + canonical_solution concatenated into a runnable module.

Provenance and licensing

Both files are derived from HumanEval (https://github.com/openai/human-eval), released by OpenAI under the MIT License. canonicals.jsonl reproduces HumanEval's original prompts and reference solutions verbatim (field-renamed for our pipeline). spec_pairs.jsonl is new work built on top of HumanEval: for each of the 49 selected HumanEval tasks, we authored an ambiguous/gold prompt pair and a decomposed set of constraint tests by hand, using the HumanEval problem statement and reference solution as the source of ground truth. task_id values (e.g. "HumanEval/26") directly reference the corresponding upstream HumanEval problem.

This derived data is released under the same MIT License as HumanEval, in accordance with its terms.