Add dataset loader

openexempt.py

@@ -0,0 +1,165 @@
+import json
+import datasets
+from pathlib import Path
+
+
+_DESCRIPTION = "OpenExempt is a diagnostic benchmark for legal reasoning in language models."
+_HOMEPAGE = "https://github.com/servantez/OpenExempt"
+_LICENSE = "CC BY 4.0"
+_VERSION = datasets.Version("1.0.0")
+_CITATION = """"""
+
+_SUITES = {
+    "advanced_competency": {
+        "archive": "data/advanced_competency.tar.gz",
+        "description": "Advanced Competency Suite."
+    },
+    "basic_competency": {
+        "archive": "data/basic_competency.tar.gz",
+        "description": "Basic Competency Suite."
+    },
+    "intermediate_competency": {
+        "archive": "data/intermediate_competency.tar.gz",
+        "description": "Intermediate Competency Suite."
+    },
+    "asset_scaling": {
+        "archive": "data/asset_scaling.tar.gz",
+        "description": "Asset Scaling Suite."
+    },
+    "temporal_reasoning": {
+        "archive": "data/temporal_reasoning.tar.gz",
+        "description": "Temporal Reasoning Suite."
+    },
+    "reasoning_decomposition": {
+        "archive": "data/reasoning_decomposition.tar.gz",
+        "description": "Reasoning Decomposition Suite."
+    },
+    "baseline_robustness": {
+        "archive": "data/baseline_robustness.tar.gz",
+        "description": "Baseline Robustness Suite."
+    },
+    "distractor_robustness": {
+        "archive": "data/distractor_robustness.tar.gz",
+        "description": "Distractor Robustness Suite."
+    },
+    "obfuscation_robustness": {
+        "archive": "data/obfuscation_robustness.tar.gz",
+        "description": "Obfuscation Robustness Suite."
+    },
+    "sycophancy_robustness": {
+        "archive": "data/sycophancy_robustness.tar.gz",
+        "description": "Sycophancy Robustness Suite."
+    },
+}
+
+def read_json(path: Path):
+    with path.open("r", encoding="utf-8") as file:
+        return json.load(file)
+
+def read_jsonl_file(path: Path):
+    with path.open("r", encoding="utf-8") as file:
+        return [json.loads(line) for line in file if line.strip()]
+
+class OpenExemptConfig(datasets.BuilderConfig):
+
+    def __init__(self, suite, **kwargs):
+        if suite == "all":
+            description = _DESCRIPTION
+            archives = [info["archive"] for info in _SUITES.values()]
+        else:
+            info = _SUITES[suite]
+            description = f"OpenExempt: {info['description']}"
+            archives = [info["archive"]]
+        super(OpenExemptConfig, self).__init__(
+            name=suite, 
+            description=description,
+            version=_VERSION,
+            **kwargs)
+        self.suite = suite
+        self.archives = archives
+        
+class OpenExempt(datasets.GeneratorBasedBuilder):
+    BUILDER_CONFIG_CLASS = OpenExemptConfig
+    BUILDER_CONFIGS = [
+        OpenExemptConfig(suite=suite) 
+        for suite in ["all"] + list(_SUITES.keys())
+    ]
+    DEFAULT_CONFIG_NAME = "all"
+
+    def _info(self):
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+            homepage=_HOMEPAGE,
+            license=_LICENSE,
+            citation=_CITATION,
+            features=datasets.Features(
+                {
+                    "id": datasets.Value("string"),
+                    "prompt": datasets.Value("string"),
+                    "solution": datasets.Value("string"),
+                    "config": datasets.Value("string"),
+                    "case": datasets.Value("string")
+                }
+            ),
+        )
+
+    def _split_generators(self, dl_manager):
+        extracted_paths = dl_manager.download_and_extract(self.config.archives)
+        dataset_dirs = []
+        for extracted_path in extracted_paths:
+            suite_dir = next(Path(extracted_path).iterdir())
+            for dataset_dir in suite_dir.iterdir():
+                if dataset_dir.is_dir():
+                    dataset_dirs.append(str(dataset_dir))
+        return [
+            datasets.SplitGenerator(
+                name=datasets.Split.VALIDATION,
+                gen_kwargs={
+                    "dataset_dirs": dataset_dirs,
+                    "split": "dev",
+                },
+            ),
+            datasets.SplitGenerator(
+                name=datasets.Split.TEST,
+                gen_kwargs={
+                    "dataset_dirs": dataset_dirs,
+                    "split": "test",
+                },
+            ),
+        ]
+
+    def _generate_examples(self, dataset_dirs, split):
+        for dataset_dir in dataset_dirs:
+            dataset_dir = Path(dataset_dir)
+
+            config = read_json(dataset_dir / "config.json")
+            shared = read_json(dataset_dir / "shared.json")
+            examples = read_jsonl_file(dataset_dir / f"{split}.jsonl")
+            case_file_name = 'cases' if split == 'test' else f'{split}_cases'
+            cases = read_jsonl_file(dataset_dir / f'{case_file_name}.jsonl')
+
+            if len(examples) != len(cases):
+                raise ValueError(f"Number of examples and cases do not match for dataset: {dataset_dir}")
+            for example, case in zip(examples, cases):
+                uid = example["uid"]
+                prompt_inputs = [
+                    shared["instruction"], 
+                    shared["meta_instruction"], 
+                    shared["response_format"], 
+                    example["facts"], 
+                    example.get("solved_steps"), # Solved steps can be omitted
+                    shared["statutes"],
+                    shared["format_reminder"]
+                ]
+                prompt = '\n\n'.join(filter(None, prompt_inputs))
+                solution = example["solution"]
+                if isinstance(solution, dict):
+                    solution = json.dumps(solution, sort_keys=True)
+                
+                yield uid, {
+                    "id": uid,
+                    "prompt": prompt,
+                    "solution": solution,
+                    "config": json.dumps(config, sort_keys=True),
+                    "case": json.dumps(case, sort_keys=True)
+                }