benchmark/README.md

# APM Benchmark Scripts

This folder documents the command-line scripts used to run the Assistive Prompt Mediation (APM) benchmark from the published prompt CSVs through metric aggregation.

The scripts live in `scripts/`:

- `prepare_inference_inputs.py` converts the dataset CSVs into deterministic JSONL examples.
- `evaluate_outputs.py` computes row-level protocol, script, and burden metrics from model outputs.
- `compile_results.py` merges those metrics with judge annotations.
- `analyze_results.py` writes benchmark summary CSVs and optional plots.

## Setup

```bash
python -m pip install -r requirements-benchmark.txt
```

The row-level metric and compilation scripts only use the Python standard library. `pandas` is needed for aggregate analysis; `matplotlib` is only needed when `--plots` is used.

## 1. Prepare Inference Inputs

Run from the dataset repository root:

```bash
python scripts/prepare_inference_inputs.py \
  --dataset-dir . \
  --output-dir benchmark_inputs \
  --overwrite
```

This writes files such as `benchmark_inputs/N1/en.jsonl`. Each row has the following shape:

```json
{
  "example_id": "stable_sha1_id",
  "language": "en",
  "noise": "N1",
  "category": "Communication and Task delegation",
  "alpha": 0.2,
  "clean_text": "Write a short message asking my manager for one day of sick leave.",
  "noisy_prompt": "Wriet a shrot emssage askign my managre fro noe dya of sikc elave.",
  "instruction": "Rewrite the noisy user prompt into a clear prompt..."
}
```

The full CSV set expands to 40,928 inference examples across languages, noise types, and alpha severities.

Save model generations in this format before evaluation:

```json
{
  "example_id": "stable_sha1_id",
  "language": "en",
  "noise": "N1",
  "alpha": 0.2,
  "noisy_prompt": "Wriet a shrot emssage askign my managre fro noe dya of sikc elave.",
  "model_response": "Write a short message asking my manager for one day of sick leave."
}
```

The evaluator also accepts common response field names such as `response`, `assistant_response`, `assist_prompt`, `assisted_prompt`, `mediated_prompt`, or `output`.

## 2. Evaluate Model Outputs

Expected layout:

```text
outputs/
  model-name/
    N1/
      results.jsonl
    N2/
      results.jsonl
```

Run:

```bash
python scripts/evaluate_outputs.py \
  --input-root outputs \
  --output-root metrics
```

This writes `metrics/<model>/<noise>/metrics.json`.

Each metric row includes:

- `asked_question`, `added_explanation`, `protocol_compliant`
- `script_match`, `script_ratio` for Chinese-script diagnostics
- `B_raw`, `B_assist`, and `BRS = B_raw - B_assist`

## 3. Compile With Judge Annotations

If judge annotations are stored as `outputs_judged/<model>/<noise>/results.jsonl`, run:

```bash
python scripts/compile_results.py \
  --metrics-root metrics \
  --judged-root outputs_judged \
  --output-root compiled
```

Judge annotations may be nested under a `judge` key:

```json
{
  "example_id": "stable_sha1_id",
  "judge": {
    "intent_preservation": 5,
    "hallucinated_additions": false,
    "hallucination_severity": 0,
    "asked_for_clarification": false,
    "added_extra_text": false,
    "language_match": true,
    "overall_verdict": "pass",
    "comment": "Intent is preserved."
  }
}
```

The compiler prefixes nested judge fields with `judge_` and writes `compiled/<model>/<noise>/compiled.json`.

## 4. Aggregate Benchmark Results

```bash
python scripts/analyze_results.py \
  --compiled-root compiled \
  --output-dir benchmark_summaries
```

Add `--plots` to also write PNG visualizations when matplotlib is available in your environment.

Summary outputs include:

- `apm_sensitivity_curves.csv`
- `core_metrics.csv`
- `hallucination_metrics.csv`
- `intent_burden_tradeoff.csv`
- `false_robustness_summary.csv`
- `language_noise_disparities.csv`
- `table_language_avg.csv`
- `table_model_avg.csv`

Use `--write-combined` if you also want a flattened `compiled_results.csv`.