Add dataset configuration to fix schema mismatch

README.md

@@ -0,0 +1,209 @@
+---
+configs:
+  # Main benchmark dataset (raw code pairs with timing)
+  - config_name: benchmark
+    data_files:
+      - split: cpp
+        path: final_benchmark_uniform/cpp_seed42_sanitized.jsonl
+      - split: python
+        path: final_benchmark_uniform/python_seed42_sanitized.jsonl
+      - split: java
+        path: final_benchmark_uniform/java_seed42_sanitized_process.jsonl
+
+  # Zero-shot format (system_prompt + user_prompt + code_a + code_b + expected)
+  - config_name: zeroshot
+    data_files:
+      - split: cpp_test
+        path: cpp_zeroshot_test.jsonl
+      - split: python_test
+        path: python_zeroshot_test.jsonl
+      - split: java_test
+        path: java_zeroshot_test.jsonl
+
+  # A/B format (instruction + input + output) - for instruction tuning
+  - config_name: ab
+    data_files:
+      - split: cpp_train
+        path: cpp_ab_train.jsonl
+      - split: cpp_val
+        path: cpp_ab_val.jsonl
+      - split: cpp_test
+        path: cpp_ab_test.jsonl
+      - split: python_train
+        path: python_ab_train.jsonl
+      - split: python_val
+        path: python_ab_val.jsonl
+      - split: python_test
+        path: python_ab_test.jsonl
+      - split: java_train
+        path: java_ab_train.jsonl
+      - split: java_val
+        path: java_ab_val.jsonl
+      - split: java_test
+        path: java_ab_test.jsonl
+
+  # A/B with diff format
+  - config_name: abdiff
+    data_files:
+      - split: cpp_train
+        path: cpp_abdiff_train.jsonl
+      - split: cpp_val
+        path: cpp_abdiff_val.jsonl
+      - split: cpp_test
+        path: cpp_abdiff_test.jsonl
+      - split: python_train
+        path: python_abdiff_train.jsonl
+      - split: python_val
+        path: python_abdiff_val.jsonl
+      - split: python_test
+        path: python_abdiff_test.jsonl
+      - split: java_train
+        path: java_abdiff_train.jsonl
+      - split: java_val
+        path: java_abdiff_val.jsonl
+      - split: java_test
+        path: java_abdiff_test.jsonl
+
+  # A with diff format
+  - config_name: adiff
+    data_files:
+      - split: cpp_train
+        path: cpp_adiff_train.jsonl
+      - split: cpp_val
+        path: cpp_adiff_val.jsonl
+      - split: cpp_test
+        path: cpp_adiff_test.jsonl
+      - split: python_train
+        path: python_adiff_train.jsonl
+      - split: python_val
+        path: python_adiff_val.jsonl
+      - split: python_test
+        path: python_adiff_test.jsonl
+      - split: java_train
+        path: java_adiff_train.jsonl
+      - split: java_val
+        path: java_adiff_val.jsonl
+      - split: java_test
+        path: java_adiff_test.jsonl
+
+license: mit
+task_categories:
+  - text-classification
+language:
+  - en
+  - code
+tags:
+  - code
+  - efficiency
+  - benchmark
+  - llm-evaluation
+  - code-optimization
+size_categories:
+  - 1K<n<10K
+---
+
+# EffiBench Dataset
+
+A comprehensive benchmark for evaluating Large Language Models' ability to judge code execution efficiency.
+
+## 📊 Dataset Overview
+
+This dataset contains pairs of functionally equivalent code snippets with different execution speeds, designed to evaluate whether LLMs can correctly identify which code runs faster.
+
+- **Languages**: C++, Python, Java
+- **Samples**: 900 pairs per language (2,700 total)
+- **Difficulty Levels**: Hard (2-4x speedup), Medium (4-8x), Easy (8x+)
+
+## 📁 Configurations
+
+| Config | Description | Splits | Schema |
+|--------|-------------|--------|--------|
+| `benchmark` | Main benchmark (raw code pairs) | cpp, python, java | `slow_code`, `fast_code`, `speedup`, `difficulty` |
+| `zeroshot` | Zero-shot LLM evaluation | {lang}_test | `system_prompt`, `user_prompt`, `code_a`, `code_b`, `expected` |
+| `ab` | A/B comparison (instruction tuning) | {lang}_train/val/test | `instruction`, `input`, `output` |
+| `abdiff` | A/B with code diff | {lang}_train/val/test | `instruction`, `input`, `output` |
+| `adiff` | Single code with diff | {lang}_train/val/test | `instruction`, `input`, `output` |
+
+## 🚀 Usage
+
+```python
+from datasets import load_dataset
+
+# Load main benchmark dataset
+ds = load_dataset("JinNian0072/efficodebench-dataset", "benchmark")
+print(ds["cpp"][0])  # First C++ sample
+
+# Load zero-shot evaluation format
+ds_zs = load_dataset("JinNian0072/efficodebench-dataset", "zeroshot")
+print(ds_zs["python_test"][0])
+
+# Load A/B format for training
+ds_ab = load_dataset("JinNian0072/efficodebench-dataset", "ab")
+print(ds_ab["cpp_train"][0])
+
+# Load specific language and split
+ds_java = load_dataset("JinNian0072/efficodebench-dataset", "abdiff", split="java_test")
+```
+
+## 📋 Data Format
+
+### Benchmark Config
+```json
+{
+  "pair_id": "pie_12345",
+  "problem_id": "p01234",
+  "language": "cpp",
+  "slow_code": "...",
+  "fast_code": "...",
+  "slow_time": 1234.5,
+  "fast_time": 123.4,
+  "speedup": 10.0,
+  "difficulty": "easy"
+}
+```
+
+### Instruction Tuning Config (ab/abdiff/adiff)
+```json
+{
+  "instruction": "Which code is more efficient?",
+  "input": "Code A: ...\nCode B: ...",
+  "output": "A",
+  "pair_id": "pie_12345",
+  "language": "cpp"
+}
+```
+
+### Zero-shot Config
+```json
+{
+  "system_prompt": "You are a code efficiency expert...",
+  "user_prompt": "Compare the following code snippets...",
+  "code_a": "...",
+  "code_b": "...",
+  "expected": "A",
+  "difficulty": "hard"
+}
+```
+
+## 📈 Difficulty Levels
+
+| Difficulty | Speedup Range | Description |
+|------------|---------------|-------------|
+| hard       | [2x, 4x)      | Small optimization margin |
+| medium     | [4x, 8x)      | Moderate optimization |
+| easy       | [8x, inf)     | Large optimization margin |
+
+## 📄 License
+
+MIT License
+
+## 📬 Citation
+
+```bibtex
+@misc{efficodebench2026,
+  title={EffiBench: Evaluating LLMs on Code Efficiency Judgment},
+  author={...},
+  year={2026},
+  url={https://huggingface.co/datasets/JinNian0072/efficodebench-dataset}
+}
+```