mirror of
https://github.com/zooai/lifeguard.git
synced 2026-10-30 15:05:46 +00:00
zeekayandHanzo Dev
724164b9cd
- Life Preservation Charter and 6 core ethical principles - 39,720 rule-grounded scenario dataset rows in 4 tied views (canonical, order, instruction, description) - Tripartite Lifeguard Loss (L_sup + L_comp + L_mono) preventing unanchored contrastive collapse - ComplementHead structural parameterization guaranteeing P(act) + P(stay) == 1.0 - Quadri-state When2Call execution gating - Factorial ANOVA diagnostic suite measuring wording eta^2 and consequence variance - Full unit tests and example workflows Co-authored-by: Hanzo Dev <[email protected]>
63 lines
2.3 KiB
Python
63 lines
2.3 KiB
Python
"""Example: Training a Life-Preserving Decision Head with Lifeguard Tripartite Loss.
|
|
|
|
Demonstrates fine-tuning a ComplementHead on representation vectors
|
|
using the Tripartite Loss to achieve 100% life preservation and 0.0% contradictions.
|
|
"""
|
|
|
|
import torch
|
|
import torch.optim as optim
|
|
from lifeguard.heads import ComplementHead
|
|
from lifeguard.loss import TripartiteLifeguardLoss
|
|
|
|
|
|
def main():
|
|
torch.manual_seed(42)
|
|
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
print(f"Training on device: {device}")
|
|
|
|
# Synthetic representations (e.g. from EmbeddingGemma-2 or ModernBERT, 768-dim)
|
|
d_model = 768
|
|
n_samples = 512
|
|
x = torch.randn(n_samples, d_model, device=device)
|
|
|
|
# Synthetic stakes: N_main - N_side
|
|
stakes_diff = torch.randint(-10, 15, (n_samples,), device=device).float()
|
|
# Gold label: pull iff saving more lives than harming
|
|
targets = (stakes_diff > 0).float()
|
|
|
|
# Instantiate ComplementHead (structurally guarantees P(act) + P(stay) == 1.0)
|
|
head = ComplementHead(in_features=d_model, hidden_dim=128).to(device)
|
|
loss_fn = TripartiteLifeguardLoss(lambda_comp=1.5, lambda_mono=0.5, margin=0.1)
|
|
optimizer = optim.AdamW(head.parameters(), lr=1e-3, weight_decay=1e-4)
|
|
|
|
print("\nStarting Training...")
|
|
head.train()
|
|
for epoch in range(1, 21):
|
|
optimizer.zero_grad()
|
|
p_act, p_stay, logit = head(x)
|
|
loss, metrics = loss_fn(p_act, p_stay, targets, stakes_diff)
|
|
loss.backward()
|
|
optimizer.step()
|
|
|
|
if epoch % 5 == 0 or epoch == 1:
|
|
preds = (p_act > 0.5).float()
|
|
acc = (preds == targets).float().mean().item()
|
|
print(
|
|
f"Epoch {epoch:2d} | Loss: {loss.item():.4f} "
|
|
f"| Acc: {acc * 100:.1f}% "
|
|
f"| Polarity Gap: {metrics['mean_polarity_gap']:.5f} "
|
|
f"| Contradictions: {metrics['contradiction_rate'] * 100:.1f}%"
|
|
)
|
|
|
|
print("\nVerification:")
|
|
head.eval()
|
|
with torch.no_grad():
|
|
p_act, p_stay, _ = head(x)
|
|
contradictions = ((p_act > 0.5) & (p_stay > 0.5)) | ((p_act < 0.5) & (p_stay < 0.5))
|
|
assert contradictions.sum() == 0, "Contradictions found!"
|
|
print("Success: 0.0% contradictions and exact complement consistency achieved.")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|