File size: 1,883 Bytes
f8c0823 | 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 | """Revisión independiente de etiquetas (Claude Opus 5.5, no humana) sobre una muestra estratificada.
S = claramente discriminatorio sin contexto adicional; A = ambiguo o depende del contexto; N = no discriminatorio."""
import json
import pandas as pd
def ids(prefix, nums): return {f"{prefix}-{n:02d}" for n in nums}
J = {}
J.update({i: "A" for i in ids("rac-1", [2, 10, 12, 16, 35])})
J.update({i: "N" for i in ids("cla-1", [0, 7, 30, 39])}); J.update({i: "A" for i in ids("cla-1", [2, 4, 17, 19, 25, 31])})
J.update({i: "S" for i in ids("mac-1", [0, 2, 3, 6, 7, 8, 11, 17, 19, 21, 22, 25, 27, 33, 34, 37, 39])})
J.update({i: "N" for i in ids("mac-1", [15, 16, 20, 24, 26, 28, 30, 31, 32, 38])})
J.update({i: "S" for i in ids("sin-1", [16, 25, 44])})
J.update({i: "A" for i in ids("sin-1", [0, 2, 9, 10, 13, 15, 18, 21, 22, 29, 33, 38, 40, 41, 45])})
df = pd.read_csv("muestra_revision.csv")
def judge(r):
if r.id in J: return J[r.id]
if r.etiqueta == 0: return "N" # contrastes: todos revisados como no discriminatorios
return {"rac": "S", "cla": "S", "mac": "A", "sin": "N"}[r.ds] # resto por defecto según la revisión
df["juicio"] = df.apply(judge, axis=1)
df["revisor"] = "Claude Opus 5.5 (revisión de modelo, no humana)"
df.to_csv("revision_etiquetas.csv", index=False)
res = {}
for ds, g in df.groupby("ds"):
p, n = g[g.etiqueta == 1], g[g.etiqueta == 0]
res[ds] = {"n_discriminatorios": len(p), "claros_pct": round(100 * (p.juicio == "S").mean(), 1),
"ambiguos_pct": round(100 * (p.juicio == "A").mean(), 1), "no_discriminatorios_pct": round(100 * (p.juicio == "N").mean(), 1),
"n_contraste": len(n), "contraste_correcto_pct": round(100 * (n.juicio == "N").mean(), 1)}
json.dump(res, open("resultados_revision.json", "w"), ensure_ascii=False, indent=2)
print(json.dumps(res, ensure_ascii=False, indent=1))
|