| """ |
| Shared resolution of per-rater synthetic-ratings tables. |
| |
| Every analysis that consumes synthetic ratings takes a `--model <slug>` and |
| routes through here, so adding a rater never means editing a hardcoded path. |
| |
| data/main_study/synth_long_{audio,text}_table_<slug>.csv |
| |
| `gpt4o` falls back to the unsuffixed tables the paper was written against, so |
| running any of these scripts with no arguments reproduces the published |
| numbers exactly. |
| """ |
|
|
| from __future__ import annotations |
|
|
| from pathlib import Path |
|
|
| import pandas as pd |
|
|
| STUDY = Path("data/main_study") |
| DEFAULT_MODEL = "gpt4o" |
|
|
|
|
| def synth_path(model: str, modality: str) -> Path: |
| """Path to one rater's long table. modality in {'audio', 'text'}.""" |
| path = STUDY / f"synth_long_{modality}_table_{model}.csv" |
| if path.exists(): |
| return path |
|
|
| fallback = STUDY / f"synth_long_{modality}_table.csv" |
| if model == DEFAULT_MODEL and fallback.exists(): |
| return fallback |
|
|
| raise FileNotFoundError( |
| f"{path} not found. Generate it with:\n" |
| f" python analysis/01_data_prep/03_synth_to_long_csv_multi.py " |
| f"--model {model}" |
| ) |
|
|
|
|
| def load_synth(model: str, modality: str) -> pd.DataFrame: |
| return pd.read_csv(synth_path(model, modality)) |
|
|
|
|
| def add_model_arg(parser) -> None: |
| parser.add_argument( |
| "--model", default=DEFAULT_MODEL, |
| help=f"synthetic rater slug (default: {DEFAULT_MODEL}, the published run)", |
| ) |
|
|
|
|
| def tag(model: str) -> str: |
| """ |
| Filename suffix for outputs. |
| |
| Empty for the default rater so the published artifacts keep their names and |
| are overwritten in place; '_<slug>' for anything else so a new rater cannot |
| silently clobber the paper's figures. |
| """ |
| return "" if model == DEFAULT_MODEL else f"_{model}" |
|
|
