tasks/candidate-0674-ml-evaluation/instruction.md

You are an agent, your current working directory is /app.

You can use the tools available to you to interact with the computer to assist the user in completing tasks.

Task

Repair the frozen TorchMetrics retrieval source tree under /app/vendor/torchmetrics so all four public evaluation paths enforce one consistent top_k contract:

• retrieval_average_precision
• retrieval_reciprocal_rank
• RetrievalMAP
• RetrievalMRR

Required behavior

1. top_k=None evaluates the complete ranking for each query.
2. A positive integer evaluates only the highest-scored k documents within each query.
3. Zero, negative integers, non-integral numbers, and other non-integer values must raise ValueError at the public API boundary.
4. Ranking remains descending by prediction score, preserving each score's relevance label.
5. Stateful metrics must group by query index before truncation and preserve empty_target_action plus mean, median, min, and max aggregation behavior.
6. Functional and stateful APIs must agree on valid inputs.

Modify the existing production modules rather than adding a replacement evaluator. The intended repair spans the two functional retrieval modules and the two stateful retrieval modules. Do not delete or rewrite the supplied source tree, public cases, or workflow helper.

After repairing the modules, run:

python3 /app/tools/run_retrieval_workflow.py

The command must finish successfully and write /app/evaluation_report.json with schema version retrieval_topk_eval.v1. Runtime networking and package installation are not allowed.