hadro/cde-evals

City Directory Extraction — Gold Evaluation Sets

Evaluation only. Never train on these. This is the regression harness for hadro/city-directory-extraction — the gold sets used to score every fine-tune of the Qwen3.5 city-directory extractor.

Why this repo is private. Two reasons, and the second is the stronger one:

1. nyu_eval.jsonl is CC-BY-SA-NC 4.0 (non-commercial + share-alike).
2. Public gold sets get scraped into training corpora. If this panel leaks into a pretraining crawl, it stops measuring anything. Keep it out of public mirrors, public HF Spaces, and pasted-into-a-chat contexts.

Schema

Every file is JSONL, one directory entry per line:

{
  "raw_line": "Brower N. merchant, 95, Water-street",
  "context": {"publisher": "franks", "alphabetical_range": "", "directory_year": "1786", "image": "..."},
  "record": {
    "name": "Brower N.", "is_business": false, "spouse_name": "", "race_designation": "",
    "occupation_role": "merchant", "employer": "", "address": "95 Water-street", "home_address": ""
  }
}

• raw_line — the entry as printed (model input).
• context — page-level metadata fed in the prompt, not predicted.
• record — the 8-field target. context.image is present on the panel volumes only.

The 18-volume NYC panel — the regression harness

Hand-labeled by this project from library page scans (NYPL / Internet Archive / LoC), one gold set per volume, all validator-clean. NYC 1786–1933 across boroughs and publishers. 1169 lines total. This is the panel, not a sample — score all 18.

File	Rows
franks1786_eval.jsonl	56
duncan1794_eval.jsonl	58
mercein1820_eval.jsonl	60
ogden1839_eval.jsonl	66
doggett1846_eval.jsonl	37
rode1851_eval.jsonl	53
hearne1852_eval.jsonl	52
hopehenderson1856_eval.jsonl	60
lain1876_eval.jsonl	103
boyd1890_eval.jsonl	75
trow1907_eval.jsonl	68
trow1913_eval.jsonl	93
polk1917_eval.jsonl	72
polk1925_eval.jsonl	40
mb1931_eval.jsonl	109
polk1933bk_eval.jsonl	49
polk1933si_eval.jsonl	56
queens1933_eval.jsonl	62

Externals + controls

File	Rows	Source	License	Notes
synth_dev.jsonl	1,000	generated by synth_persons.py	ours (CC-BY-4.0)	In-distribution control. Near-ceiling scores are expected; it detects breakage, not quality.
nyu_eval.jsonl	3,000	NYU NYC directories 1850–1890	CC-BY-SA-NC 4.0 ⚠	Score at --limit 500 — every published baseline is on the first 500 rows.
ftd_eval.jsonl	2,000	French Trade Directories (SODUCO)	open (CC)	Cross-lingual transfer eval.
tulsa_eval.jsonl	1,500	own directory-pipeline output (Tulsa 1921)	ours	Real OCR + Gemini-extracted (silver) labels, not hand-verified.
lain_eval.jsonl	800	own pipeline output (Lain Brooklyn 1897)	ours	Silver labels. Held out of the sampling/harvest set.
minneapolis_eval.jsonl	230	DirCity	MIT	Out-of-city US eval (silver).

Note lain_eval.jsonl (silver, 800 rows, full volume) and lain1876_eval.jsonl (gold, 103 rows, panel volume) are different sets — don't merge them.

How to run it

The harness lives in the public repo; only the data is here.

git clone https://github.com/hadro/city-directory-extraction
cd city-directory-extraction
hf download hadro/cde-evals --repo-type dataset --local-dir data/

Then, on the cluster:

source hpc/env.sh
sbatch $(slurm_gpu_args) --export=ALL,RUN_NAME=<run> hpc/30_eval.sbatch

hpc/30_eval.sbatch runs the 18 panel volumes, then the externals, then eval/evaluate.py and the eval/results_table.py pivot. It sets HF_HUB_OFFLINE=1, so download this dataset on the login node first — compute nodes have no network.

Two gotchas that have each cost a full eval cycle:

• Pass --base-model. train/sft_qwen.py trains LoRA, so the run directory is an adapter, not a full checkpoint.
• Watch the log for missing adapter keys. That warning means the adapter silently did not apply and you scored the bare base model. qwen_predict.py must load AutoModelForImageTextToText (not AutoModelForCausalLM) — Qwen3.5 is multimodal and all-linear adapts the vision tower too. This bug produced a 0.358 macro-F1 that was really 0.760.

Scoring: macro-F1 averages over the fields the gold actually has; micro-F1 is frequency-weighted. They disagree, and both are reported. See eval/evaluate.py.

Provenance and redistribution

Mixed licensing — there is no single license covering this repo:

• The 18-volume panel is this project's own hand labeling of public-domain page scans.
• nyu_eval.jsonl is CC-BY-SA-NC 4.0. Non-commercial and share-alike. It is used for evaluation only and is never mixed into training, which is what keeps the released synthetic-trained model permissively reusable.
• ftd_eval.jsonl is open CC from the SODUCO project.
• minneapolis_eval.jsonl derives from an MIT-licensed repo.
• tulsa/lain are this project's own pipeline output.

Check the terms before redistributing any of it. The silver-labeled sets (tulsa, lain, minneapolis) are machine-extracted and not hand-verified — treat their absolute numbers as approximate and use them for relative comparison across runs.

Related

• hadro/city-directory-extraction — code, harness, docs
• hadro/city-directory-synth — synthetic training data (public)
• hadro/city-dir-08b-yaml-v5 — current best adapter