probe_quality.py

"""Fit-for-purpose check: can this prompt make laya separate Ekman's anger from contempt at all?"""
import numpy as np, torch, laya_opt
from ekman_questions import EKMAN_QUESTIONS
from probes import PROBES

agent = laya_opt.load_agent()
texts=[t for _,t in PROBES]; gold=[g for g,_ in PROBES]
res, st = laya_opt.score_texts(agent, EKMAN_QUESTIONS, texts, max_len=256, head_max_len=144, token_budget=6144)
sup = [r.get("superiority",{}).get("noul") for r in res]
blk = [r.get("blocked_or_unfair",{}).get("noul") for r in res]
print("%-10s %-8s %-8s %-6s %-6s %s" % ("gold","p_anger","p_cont","conf","margin","flags"))
ok=0
for g,r,su,bl in zip(gold,res,sup,blk):
    pa,pc = r["ekman"]["probabilities"]["anger"], r["ekman"]["probabilities"]["contempt"]
    pred = "anger" if pa>=pc else "contempt"; ok += pred==g
    print("%-10s %-8.3f %-8.3f %-6.2f %-6.3f sup=%.2f blk=%.2f%s" % (g,pa,pc,r["ekman"]["confidence"],pa-pc,su,bl,"  <-- WRONG" if pred!=g else ""))
print("\nchoice-head accuracy on the 16 probes: %d/16 = %.3f  (chance = 0.50)" % (ok, ok/len(gold)))
n=int(len(gold))
sup_c = np.mean([s for g,s in zip(gold,sup) if g=="contempt"]); sup_a=np.mean([s for g,s in zip(gold,sup) if g=="anger"])
blk_a = np.mean([b for g,b in zip(gold,blk) if g=="anger"]); blk_c=np.mean([b for g,b in zip(gold,blk) if g=="contempt"])
print("noul probes: P(superiority) contempt=%.2f anger=%.2f | P(blocked/unfair) anger=%.2f contempt=%.2f" % (sup_c,sup_a,blk_a,blk_c))
print("intensity  : " + ", ".join("%.1f"%r["intensity"]["score"] for r in res))
print("timing     : %s" % st)