ztc/usage.py

# Zero-Token Confidence (ZTC) — usage
#
# ZTC reads the model's own internal state ONCE, before generation, and returns
# the probability that the answer the model is about to produce will be correct.
# No extra tokens are generated. No second model is required.
#
#   probe file : ztc/ztc_probe_darwin397b.npz   (45 KB)
#   input      : final-layer hidden state of the last prompt token (4096-dim)
#   output     : score, and a calibrated probability in [0, 1]
#
# Reported performance on this model (PubMedQA, 539 items, 146 incorrect):
#   self-reported confidence  AUROC 0.7646
#   ZTC                       AUROC 0.8801      (permutation null z = 13.31)

import numpy as np
import torch
from transformers import AutoModel, AutoTokenizer

MODEL = "FINAL-Bench/Darwin-397B-ZTC"
PROBE = "ztc/ztc_probe_darwin397b.npz"


class ZTC:
    def __init__(self, path=PROBE):
        z = np.load(path)
        self.w = z["w"].astype(np.float32)
        self.mu = z["mu"].astype(np.float32)
        self.sd = z["sd"].astype(np.float32)
        self.s_mean = float(z["s_mean"]); self.s_std = float(z["s_std"])
        self.A = float(z["cal_A"]); self.B = float(z["cal_B"])

    def score(self, hidden):
        """hidden: (4096,) or (batch, 4096) final-layer state of the last prompt token."""
        h = np.asarray(hidden, dtype=np.float32)
        s = ((h - self.mu) / self.sd) @ self.w
        p = 1.0 / (1.0 + np.exp(-(self.A * (s - self.s_mean) / self.s_std + self.B)))
        return s, p


# --- one forward pass, zero generated tokens -------------------------------
tok = AutoTokenizer.from_pretrained(MODEL)
model = AutoModel.from_pretrained(MODEL, dtype=torch.bfloat16, device_map="auto").eval()
ztc = ZTC()

prompt = "Question: ...\nAnswer:"
b = tok(prompt, return_tensors="pt").to(next(model.parameters()).device)
with torch.no_grad():
    h = model(**b).last_hidden_state[0, -1].float().cpu().numpy()

s, p = ztc.score(h)
print("ZTC score %.3f  ->  P(correct) = %.3f" % (s, p))

# Gate the action, not the answer:
#   if p < THRESHOLD:  do not call the tool / escalate / answer "I don't know"
#   else:              generate as usual