suture / README.md
caiotheodoro's picture
gold splits: benchmark 777, val, train_dedhi, seed-7
ece7b78 verified
|
Raw
History Blame Contribute Delete
2.32 kB
metadata
license: apache-2.0
task_categories:
  - image-text-to-text
language:
  - en
pretty_name: Suture
tags:
  - insurance
  - document-qc
  - policy-issuance
  - synthetic
  - verifier-as-oracle
configs:
  - config_name: benchmark
    data_files: data/benchmark.jsonl
  - config_name: validation
    data_files: data/validation.jsonl
  - config_name: train
    data_files: data/train.jsonl
  - config_name: train_seed7
    data_files: data/train_seed7.jsonl

Suture

Gold tasks for policy-issuance QC: a structured underwriting binder, a structured issued policy, and the exact discrepancy set. Images are not stored. Re-render with suture_forge.generate.render_doc from the GitHub repo.

Paper and adapter: caiotheodoro/suture · caiotheodoro/suture-8b. Predictions on this gold: caiotheodoro/suture-evals.

Splits

Config Seed n What
benchmark 777 1000 Contracted held-out set. Never in train.
validation 7 holdout 80 Val. Optimistic vs 777; DEDUCTIBLE_HIGHER gold n=0.
train 7+11+13+17+19+23 5120 Published SFT mix (train_dedhi).
train_seed7 7 320 Official stratified split, no class boost.

Leakprobe vs val and vs benchmark is 0. Signature = SHA-256 over sorted ground-truth fields. Build refuses overlap > 0.

Row schema

Each line is a Task: task_id, seed, binder, policy, image_*_sha256, expected (PASS/FLAG + discrepancies), signature, difficulty. Taxonomy and weights: repo CONTRACTS.md (13 classes).

Render

from datasets import load_dataset
from suture_forge.schema import Binder, Policy
from suture_forge.generate import render_doc

row = load_dataset("caiotheodoro/suture", "benchmark", split="train")[0]
png = render_doc(Binder.model_validate(row["binder"]), "BINDER",
                 row["difficulty"], row["task_id"])

Linux vs macOS raster can differ (SHA mismatch). Labels do not. Eval in the project re-renders on the GPU host.

License / use

Apache-2.0. Synthetic ACORD-style pages, not real submissions. Form numbers are public ISO ids. Do not treat this as a carrier document corpus.