Add Astro2Method benchmark data and dataset card

README.md

@@ -1,3 +1,196 @@
 ---
 license: cc-by-4.0
+language:
+- en
+task_categories:
+- question-answering
+tags:
+- astronomy
+- astrophysics
+- machine-learning
+- method-selection
+- benchmark
+size_categories:
+- n<1K
+configs:
+- config_name: default
+  data_files:
+  - split: test
+    path: benchmark.json
 ---
+
+# Astro2Method
+
+**A benchmark for machine learning method selection in astronomy.**
+
+Astro2Method contains 500 expert-reviewed binary-choice questions from 346 astronomy papers. Each question describes an astronomical problem and presents two candidate machine learning methods, architectures, variants, or pipelines. The task is to select the method supported by the comparison in the source paper.
+
+> **Evaluation notice:** In this release, every record has `correct_option: "A"`. This is the original storage order, not an evaluation-ready answer distribution. **Randomize the option order and update the answer labels before evaluation.** Always predicting A on the unmodified file scores 100% and does not measure method-selection ability. A reproducible randomization example is provided below.
+
+## Dataset at a glance
+
+| Property | Value |
+| --- | --- |
+| Questions | 500 |
+| Unique source papers | 346 |
+| Easy questions | 376 |
+| Hard questions | 124 |
+| Source-paper publication years | 2024–2026 |
+| Language | English |
+| Format | UTF-8 JSON array |
+| Evaluation split | `test` (all 500 questions) |
+| License | CC BY 4.0 |
+
+The file is a single evaluation collection; no separate training or validation set is provided. The Hugging Face dataset configuration exposes it as `test`.
+
+## Files
+
+- `benchmark.json`: the original 500 questions, two options, answer labels, difficulty labels, and source-paper identifiers.
+- `README.md`: this dataset card and usage instructions.
+
+This release contains the benchmark data. Source-paper PDFs, manuscript archives, historical retrieval corpora, masked pipeline descriptions, model predictions, and experiment code are not included.
+
+## Task and construction
+
+The accompanying manuscript is titled *Astro2Method: Benchmarking LLMs on Machine Learning Method Selection in Astronomy*.
+
+The construction process screened 1,028 astronomy papers retrieved through NASA's Astrophysics Data System (ADS), focusing on neural-network-related studies in major astronomy journals. Comparisons were extracted with LLM assistance and reviewed by human experts. A source paper can contribute more than one question.
+
+Candidate comparisons were selected for clear performance differences and supporting explanations in the source papers. Expert review covered the scientific setup, comparison evidence, question wording, answer, and difficulty label.
+
+- **Easy:** the source paper provides a direct and complete explanation of the performance difference.
+- **Hard:** the source paper provides a partial explanation of the performance difference.
+
+Comparisons without sufficient explanatory support were excluded. Difficulty describes the completeness of the source authors' explanation; it is not assigned from model accuracy. The correct answer is specific to the source study's setting and does not establish that one method is universally superior.
+
+## Schema
+
+Each record has seven string fields:
+
+| Field | Description |
+| --- | --- |
+| `question_id` | Unique question identifier, such as `Q0001`. |
+| `paper_id` | NASA ADS bibliographic code identifying the source paper. |
+| `question` | Astronomical problem or method-selection question. |
+| `option_a` | First candidate method or pipeline. |
+| `option_b` | Second candidate method or pipeline. |
+| `correct_option` | Correct option label. All values are `A` in the original file; update after shuffling. |
+| `difficulty` | `easy` or `hard`. |
+
+Example, copied from the release:
+
+```json
+{
+  "question_id": "Q0001",
+  "paper_id": "2024MNRAS.535.1844L",
+  "question": "How can one obtain accurate photometric redshift estimates for high-redshift galaxies using only flux measurements from different observational filters as input?",
+  "option_a": "LSTM",
+  "option_b": "MLP",
+  "correct_option": "A",
+  "difficulty": "hard"
+}
+```
+
+## Load the data
+
+With Python's standard library, from the repository root:
+
+```python
+import json
+from pathlib import Path
+
+records = json.loads(Path("benchmark.json").read_text(encoding="utf-8"))
+print(len(records))  # 500
+```
+
+With Hugging Face Datasets (`pip install datasets`):
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset("Xingzhongfan/Astro2Method", split="test")
+print(dataset[0])
+```
+
+If the repository is private, authenticate with a Hugging Face account that has access before downloading.
+
+## Prepare randomized evaluation inputs
+
+The following example independently swaps the two options with probability 0.5, updates each answer label, and leaves the original data unchanged. Reuse the generated order across all models being compared. The seed below is an example, not the original manuscript's evaluation permutation; independent swaps need not yield an exactly balanced A/B distribution.
+
+```python
+import json
+import random
+from pathlib import Path
+
+records = json.loads(Path("benchmark.json").read_text(encoding="utf-8"))
+rng = random.Random(42)
+evaluation_records = []
+
+for original in records:
+    item = original.copy()
+    if rng.random() < 0.5:
+        item["option_a"], item["option_b"] = item["option_b"], item["option_a"]
+        item["correct_option"] = {"A": "B", "B": "A"}[item["correct_option"]]
+    evaluation_records.append(item)
+
+# Keep labels and provenance separate from model inputs.
+prompts = {
+    item["question_id"]: (
+        "Choose the better machine learning method for the astronomical "
+        "problem. Return A or B, with a brief justification.\n\n"
+        f"Question: {item['question']}\n"
+        f"A. {item['option_a']}\n"
+        f"B. {item['option_b']}"
+    )
+    for item in evaluation_records
+}
+answer_key = {
+    item["question_id"]: item["correct_option"]
+    for item in evaluation_records
+}
+```
+
+Do not include `correct_option` in model inputs. For direct prompting, provide only the question and randomized options; keep `paper_id` and `difficulty` as evaluation metadata.
+
+## Evaluation
+
+Report **Overall**, **Easy**, and **Hard** accuracy as the number of correct predictions divided by the number of questions in the corresponding set. Uniform random guessing has expected accuracy 50%. Predefine an answer-parsing rule; count missing or invalid answers as incorrect rather than dropping them.
+
+Given `evaluation_records` from the previous example and a mapping `predictions` from question IDs to parsed labels (`"A"` or `"B"`):
+
+```python
+def accuracy(rows, predictions):
+    correct = sum(
+        predictions.get(row["question_id"]) == row["correct_option"]
+        for row in rows
+    )
+    return correct / len(rows)
+
+for name in ("overall", "easy", "hard"):
+    subset = [
+        row for row in evaluation_records
+        if name == "overall" or row["difficulty"] == name
+    ]
+    print(f"{name}: {accuracy(subset, predictions):.2%} (n={len(subset)})")
+```
+
+Record the dataset revision, random seed and option mapping, model version, prompt, decoding parameters, answer-parsing rule, and any retrieval or tool access. These usage examples illustrate evaluation; they are not the manuscript's complete experimental harness.
+
+## Intended use and limitations
+
+- Intended for evaluating language models on astronomy-specific ML method selection. Do not train on these questions when reporting held-out benchmark performance.
+- Coverage is limited by the source-paper search and the requirement for explicit method comparisons and explanations. The dataset is not representative of every astronomy problem or ML approach.
+- Source papers are publicly available, so prior exposure in model training cannot be ruled out. Web search or retrieval may directly reveal the source comparison; report such access separately from closed-book evaluation.
+- Multiple questions can come from the same paper. If creating new splits, group by `paper_id` to avoid placing questions from the same source in different splits.
+- This compact release does not contain evidence passages, answer rationales, or full experimental context. Consult the source papers when interpreting individual comparisons.
+
+## License and attribution
+
+The benchmark is distributed under [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/), as specified in this repository's dataset metadata. Attribute Astro2Method, link to this repository, and indicate any changes when redistributing the data. Source publications retain their respective rights.
+
+When using the dataset, reference [Xingzhongfan/Astro2Method](https://huggingface.co/datasets/Xingzhongfan/Astro2Method) and record the dataset revision used. Formal paper citation details are not included in this release.
+
+## Feedback
+
+Report suspected errors through the repository's [Community discussions](https://huggingface.co/datasets/Xingzhongfan/Astro2Method/discussions), including the `question_id`, `paper_id`, and supporting evidence.