Improve dataset card: Add description, links, metadata, and sample usage for MLRC-Bench

README.md

@@ -1,3 +1,100 @@
----
-license: cc-by-4.0
----
+---
+license: cc-by-4.0
+task_categories:
+  - text-generation
+  - code-generation
+tags:
+  - agents
+  - machine-learning
+  - benchmark
+  - code-generation
+language:
+  - en
+---
+
+# MLRC-Bench: Can Language Agents Solve Machine Learning Research Challenges?
+
+MLRC-Bench is a benchmark designed to quantify how effectively language agents can tackle challenging Machine Learning (ML) Research Competitions. It focuses on open research problems that demand novel methodologies by evaluating the key steps of proposing and implementing novel research methods with rigorous protocols and objective metrics.
+
+The benchmark features a curated suite of 7 competition tasks adapted from recent Machine Learning conference competitions. Our findings reveal significant challenges for LLM agents; even the best-performing tested agent (gemini-exp-1206 under MLAB) closes only 9.3% of the gap between baseline and top human participant scores. MLRC-Bench is a dynamic benchmark, designed to grow with new ML competitions and encourage rigorous, objective evaluations of AI research capabilities.
+
+*   **Paper**: [MLRC-Bench: Can Language Agents Solve Machine Learning Research Challenges?](https://huggingface.co/papers/2504.09702)
+*   **Code**: [https://github.com/yunx-z/MLRC-Bench](https://github.com/yunx-z/MLRC-Bench)
+*   **Project Page / Leaderboard**: [https://huggingface.co/spaces/launch/MLRC_Bench](https://huggingface.co/spaces/launch/MLRC_Bench)
+
+## Sample Usage
+
+To get started with MLRC-Bench, follow these steps to set up the environment and launch an agent.
+
+### Setup MLRC-Bench Core Environment
+
+First, clone the MLRC-Bench repository and navigate into it:
+
+```bash
+git clone https://github.com/yunx-z/MLRC-Bench.git
+cd MLRC-Bench
+```
+
+Next, create and activate a conda environment named `mlab` for the benchmark's core dependencies, then install the `MLAgentBench` package:
+
+```bash
+conda create -n mlab python=3.10
+conda activate mlab
+
+# Navigate into the MLAgentBench subdirectory to install it in editable mode
+cd MLAgentBench
+pip install -e .
+pip install openai
+cd .. # Go back to the MLRC-Bench root directory
+
+# Install additional system-level dependencies for the benchmark
+bash install.sh
+```
+
+### Setup Task-Specific Environment
+
+Each competition task within MLRC-Bench has its own environment. You'll need to set up a dedicated conda environment for each task you wish to run. Replace `${TASK_NAME}` with the specific task name (e.g., `llm-merging`):
+
+```bash
+# Navigate to the task's script directory
+cd MLAgentBench/benchmarks_base/${TASK_NAME}/scripts
+
+# Create and activate a dedicated conda environment for this task
+conda env create -f environment.yml --name ${TASK_NAME}
+conda activate ${TASK_NAME}
+
+# Install core benchmark components into the task environment.
+# This ensures that MLRC-Bench and MLAgentBench are available within the task's isolated environment.
+cd ../../../.. # Navigate from 'scripts/' up to the 'MLRC-Bench/' root directory
+pip install -e . # Installs MLRC-Bench (and MLAgentBench) into the active task environment
+pip install openai # Ensure OpenAI is available in the task environment
+```
+
+*(Optional)* Some competition tasks may require setting up Kaggle API authentication (`~/.kaggle/kaggle.json`). Refer to the [Kaggle API documentation](https://www.kaggle.com/docs/api) and provide manual consent to competition rules if prompted.
+
+### Launching an Agent
+
+To launch an MLAB agent on a specific task within its activated environment:
+
+```bash
+# Ensure you are in the MLRC-Bench root directory and the task's conda environment is active.
+bash launch.sh ${TASK_NAME} ${MODEL} ${GPU_ID}
+```
+You will need to specify `MY_OPENAI_API_KEY` and `MY_AZURE_OPENAI_ENDPOINT` as environment variables for OpenAI models. Supported models are listed in `MLAgentBench/LLM.py` within the repository.
+
+## Tasks
+
+The first release of MLRC-Bench includes 7 tasks adapted from recent Machine Learning conference competitions. Each task is represented as a folder in `MLAgentBench/benchmarks_base/` within the repository. Within each task folder, the `env/` directory contains files that the research agent will see at the beginning, while the `script/` folder contains additional hidden files such as `prepare.py` for downloading data.
+
+## Citation
+
+If you use MLRC-Bench in your research, please cite the following paper:
+
+```bibtex
+@article{zhang2025mlrcbench,
+  title={MLRC-Bench: Can Language Agents Solve Machine Learning Research Challenges?},
+  author={Zhang, Yunxiang and Khalifa, Muhammad and Bhushan, Shitanshu and Murphy, Grant D and Logeswaran, Lajanugen and Kim, Jaekyeom and Lee, Moontae and Lee, Honglak and Wang, Lu},
+  journal={arXiv preprint arXiv:2504.09702},
+  year={2025}
+}
+```