competition-style-ml-engineering

# Skill: competition-style-ml-engineering ## 1. Capability Definition & Real Case * **Professional Definition**: The ability to solve benchmarked ML engineering tasks in competition-like settings that require working from datasets, descriptions, grading code, and submission formats to produce valid, high-scoring outputs under resource and process constraints. * **Dimension Hierarchy**: Data and ML Workflow Engineering->Machine Learning Engineering->competition-style-ml-engineering ### Real Case **[Case 1]** * **Initial Environment**: An offline competition workspace contains a description, train.csv, test.csv, sample_submission.csv, grading code, and baseline utilities. The task has no complete working solution yet. * **Real Question**: Produce a valid submission that performs competitively on the competition metric using only the provided resources. * **Real Trajectory**: Read the competition overview and data schema, inspect the submission format, implement preprocessing and a baseline model, validate the submission locally, iterate on data handling and model choices, and generate the final CSV file. * **Real Answer**: A valid submission file is produced in the expected format and scores better than naive baselines under the local grader. * **Why this demonstrates the capability**: This tests competition-style ML engineering because the agent must combine dataset understanding, code writing, submission validity, and iterative optimization inside a leaderboard-like workflow rather than merely patching an existing training script. --- **[Case 2]** * **Initial Environment**: A more complex competition workspace includes multimodal data, longer training cycles, and stricter formatting rules. Local validation can confirm file validity but not reveal the full score directly. * **Real Question**: Engineer a robust submission pipeline that can train, evaluate, and export predictions in the required competition format. * **Real Trajectory**: Map data modalities and labels, build a training script, generate validation estimates internally, check submission validity with the local tool, debug formatting or shape errors, and refine the strongest model within runtime constraints. * **Real Answer**: The pipeline yields a valid competition submission and demonstrates sound engineering choices for improving performance under limited feedback. * **Why this demonstrates the capability**: The capability lies in handling the entire competition workflow: datasets, local validation, formatting, experiment planning, and robustness to hidden-score uncertainty. This goes beyond ordinary model improvement because correctness of the submission process itself is part of the task. ## Pipeline Execution Instructions To synthesize data for this capability, you must strictly follow a 3-phase pipeline. **Do not hallucinate steps.** Read the corresponding reference file for each phase sequentially: 1. **Phase 1: Environment Exploration** Read the exploration guidelines to discover raw knowledge seeds: `references/EXPLORATION.md` 2. **Phase 2: Trajectory Selection** Once Phase 1 is complete, read the selection criteria to evaluate the trajectory: `references/SELECTION.md` 3. **Phase 3: Data Synthesis** Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data: `references/SYNTHESIS.md`