Insurance sentence classifier built by weight editing (no gradient descent)
BertForSequenceClassification over bert-base-uncased, turned into an 8-label insurance sentence
classifier by an editor LLM that read and wrote the weights directly (closed-form math on weights and
on activations from a budgeted number of forward passes; no backprop, no optimizers). See the project
README for the protocol and results.
Labels: coverage_grant, exclusion, premium_payment, deductible_limits, claims_procedure, definitions, term_cancellation_renewal, conditions_duties
Edited tensors (18, 6.474% of parameters changed):
bert.encoder.layer.11.attention.output.LayerNorm.bias(768 entries changed)bert.encoder.layer.11.attention.output.LayerNorm.weight(768 entries changed)bert.encoder.layer.11.attention.output.dense.bias(768 entries changed)bert.encoder.layer.11.attention.output.dense.weight(589,824 entries changed)bert.encoder.layer.11.attention.self.query.bias(768 entries changed)bert.encoder.layer.11.attention.self.query.weight(589,824 entries changed)bert.encoder.layer.11.attention.self.value.bias(768 entries changed)bert.encoder.layer.11.attention.self.value.weight(589,824 entries changed)bert.encoder.layer.11.intermediate.dense.bias(3,072 entries changed)bert.encoder.layer.11.intermediate.dense.weight(2,359,296 entries changed)bert.encoder.layer.11.output.LayerNorm.bias(768 entries changed)bert.encoder.layer.11.output.LayerNorm.weight(768 entries changed)bert.encoder.layer.11.output.dense.bias(768 entries changed)bert.encoder.layer.11.output.dense.weight(2,359,296 entries changed)bert.pooler.dense.bias(768 entries changed)bert.pooler.dense.weight(589,824 entries changed)classifier.bias(0 entries changed)classifier.weight(8 entries changed)
edit_manifest.json in this folder holds the exact diff metadata (sha256 of base and edited tensors).
from transformers import pipeline
clf = pipeline("text-classification", model="sidman007/wsp-bert-insurance-v0-p50")
clf("Flood damage is not covered under this policy.")
Run: v0/p50 · held-out test accuracy: 97.30% · macro-F1: 0.9734
Source, logs and protocol: https://github.com/siddharth-ghatti/weight-space-programming-poc
Caveat: accurate but uncalibrated
This model predicts the correct label, but its softmax probabilities are flat (the winning label typically scores around 0.2 rather than 0.9+). This is a logit-scale artifact of how the editor wired the final layers: multiplying the logits by any constant changes no prediction. Use the argmax label; fit a temperature on your own held-out data before using the scores as confidences. See the project README for details and for what it would take to use this in production.