v0.1.0: Nemotron Parse receipt pilot

Status: draft. No experiment tag is published yet.

This experiment runs Nemotron-3.5-Parse through vLLM on a local RTX 4090. It preserves 24 CORD receipt inputs' provenance and 48 measured responses. All measured requests completed. One amount changed from 79,500 to 79,300. Diagnostic ROUGE-L F1 was 0.801; exact numeric-string recall was 212/241. The evaluation does not measure layout or table-structure accuracy.

Recalculate the saved experiment without a GPU:

uv sync --locked
uv run python scripts/score_receipts.py \
  --input evidence/cord-test-24 --output runs/tag-inspection/scores.json

Companion post: I ran Nemotron Parse on a 4090. Then I checked the numbers. The public post URL is pending. Do not publish this release before that post is ready. The post must link back to https://github.com/Alberto-Codes/inference-lab/releases/tag/v0.1.0.

Measured results · Reproduction

The draft-release workflow appends the exact validated public commit to the GitHub notes. Before publication, replace the pending post status with its verified public URL.