File size: 2,615 Bytes
c202fc8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 | """
Shared helpers for the synthetic-rater sampling analyses (scripts 19-21).
These live in their own module because the analysis scripts carry numeric
filename prefixes (`19_...py`), which are not importable as Python modules.
Anything more than one script needs goes here.
human_panel per prompt x rater signed preference, plus the panel gold
synth_gaps per prompt mean synthetic margin and its absolute gap
prompt_matrix per-prompt rater-preference vectors, aligned to a prompt list
"""
from __future__ import annotations
import numpy as np
def human_panel(long_df):
"""
Per prompt x rater signed preference (clip0 - clip1), plus panel gold.
Returns (wide, gold). `wide` has one row per (prompt, rater) with the
signed preference in `pref`; `gold` has one row per prompt with the
full-panel mean margin, the within-prompt rater SD, the panel size, and
the gold winner (sign of the mean margin).
"""
df = long_df[long_df["IsAttentionCheck"] == False].copy()
df["ClipIndex"] = df["AudioClip"].str.extract(r"_(\d+)\.mp3$")[0].astype(int)
wide = df.pivot_table(
index=["PromptID", "ParticipantID"],
columns="ClipIndex", values="Rating", aggfunc="first",
).reset_index()
wide.columns.name = None
wide = wide.rename(columns={0: "r0", 1: "r1"}).dropna(subset=["r0", "r1"])
wide["pref"] = wide["r0"] - wide["r1"]
gold = wide.groupby("PromptID")["pref"].agg(
gold_margin="mean", sd_within="std", n_raters="size"
).reset_index()
gold["gold_winner"] = np.sign(gold["gold_margin"])
return wide, gold
def synth_gaps(synth_df):
"""Per prompt mean synthetic margin over repeated samples, and its gap."""
s = synth_df.copy()
s["ClipIndex"] = s["AudioClip"].str.extract(r"_(\d+)\.(?:mp3|wav)$")[0].astype(int)
wide = s.pivot_table(
index=["PromptID", "synthNo"],
columns="ClipIndex", values="Rating", aggfunc="first",
).reset_index()
wide.columns.name = None
wide = wide.rename(columns={0: "s0", 1: "s1"}).dropna(subset=["s0", "s1"])
wide["margin"] = wide["s0"] - wide["s1"]
out = wide.groupby("PromptID")["margin"].agg(
synth_margin="mean", n_synth="size"
).reset_index()
out["synth_gap"] = out["synth_margin"].abs()
out["synth_winner"] = np.sign(out["synth_margin"])
return out
def prompt_matrix(panel, prompt_ids):
"""List of per-prompt rater-preference vectors, aligned to prompt_ids."""
by_prompt = {p: g["pref"].to_numpy() for p, g in panel.groupby("PromptID")}
return [by_prompt[p] for p in prompt_ids]
|