iiinova — where your digital self keeps evolving

dbtool @dbtool.bsky.social · 29/09/2026

Built for aggregate analyses (bibliometrics, demography) — not decisions about individuals. Weights CC BY-NC 4.0, inference code MIT. Need it as an API? dbtool.it If our open tools help you, support the work on GitHub Sponsors: github.com/sponsors/sheppard94g

000

dbtool @dbtool.bsky.social · 29/09/2026

Measured gains from the model card: • women: 92.98% vs 90.30% • native scripts (routed): 88.70% vs 83.84% • 14 countries above v1, 94 par, 1 below (≥400 bench rows) Where to keep v1: Togo −2.00, Chad −2.42, Samoa −4.76 pp. Full tables: huggingface.co/textpie/genderize

100

dbtool @dbtool.bsky.social · 29/09/2026

Genderize v2 is out: one open model for binary gender (M/F) + likely country (226 codes) from a personal name. Char-level, CPU-only. 94.08% on an independent 50,258-name bench (v1: 93.71%). Honest limit: weaker than v1 in Togo, Chad and Samoa. huggingface.co/textpie/genderize

100

dbtool @dbtool.bsky.social · 29/09/2026

If our open tools help you, you can support the work on GitHub Sponsors: github.com/sponsors/sheppard94g

000

dbtool @dbtool.bsky.social · 29/09/2026

On dbtool.it: /v1/participants turns a clinical abstract into the study's sex composition. On 1,351 held-out 2021-2024 RCTs it never trained on: 96.15% sex accuracy (95% CI 94.99-97.05), macro-F1 0.89. dbtool.it/participants.html

100

dbtool @dbtool.bsky.social · 29/09/2026

New: colscan, a free offline CLI. Point it at a CSV, Parquet or database and it reports, column by column, the semantic type and whether the data is personal under GDPR. Deterministic rules, nothing sent anywhere. Apache-2.0. pip install colscan → pypi.org/project/colscan

100

dbtool @dbtool.bsky.social · 20/09/2026

Worth flagging for anyone inferring gender from names at this scale: accuracy is very uneven by name origin. On our Wikidata benchmark we measure 92% on Japanese names but 64% on Chinese ones. Asymmetric error rates can propagate straight into the effect sizes.

001

dbtool @dbtool.bsky.social · 18/09/2026

genderize ultra reached 92.3% gender accuracy on Japanese public-figure names in our Wikidata benchmark, given-name first. Training overlap is unknown; this is not population accuracy. We publish limits alongside results. dbtool.it/academic.html

000

dbtool @dbtool.bsky.social · 16/09/2026

genderize cora reached 64% gender accuracy on Chinese public-figure names in our Wikidata benchmark, given-name first. Romanised names remain difficult. Training overlap is unknown; this is not population accuracy. huggingface.co/textpie/genderize

000

dbtool @dbtool.bsky.social · 14/09/2026

Our medical EBM-NLP test: 92.7% exact sample-size agreement between DeepSeek labels and expert-derived gold (114 of 123 abstracts with an unambiguous gold number). This measures the labeller, not our trained model, and is not overall PICO accuracy. dbtool.it/open-models.html

000

dbtool @dbtool.bsky.social · 13/09/2026

Build log, week 1. What's next at dbtool: one API call that turns a clinical abstract into sample size + PICO. One disk, one consumer GPU, one box. Every number ships with its test set; the public endpoint waits until it clears 80% on hand-annotated abstracts. Follow along → dbtool.it

001

dbtool @dbtool.bsky.social · 12/09/2026

genderize cora + ultra are out as open weights: byte-level, CPU-only gender + country (226 ISO codes) from a name. Gender 97.9%/98.2%, country top-1 82.6%/83.7% on 25k unseen names, network alone. CC BY-NC 4.0.

100

dbtool @dbtool.bsky.social · 12/09/2026

genderize cora & ultra are now open weights: a personal name → probable gender + a ranking of 226 countries, from 48 UTF-8 bytes. No tokenizer, no dictionary, CPU only. 98.2% gender accuracy on 25,000 held-out names. Weights CC BY-NC 4.0, script MIT. huggingface.co/textpie/genderize

000

dbtool @dbtool.bsky.social · 12/09/2026

010

dbtool @dbtool.bsky.social · 03/09/2026

Same model on every year, so the trend is free of NLM indexing artifacts. Studies covering both sexes: 32%→65%. Aged participants: 18%→39%. 88% agreement with MeSH where present. Full dataset (12.9M PMIDs) on request. @nrobinsongarcia.bsky.social may find this relevant.

011

dbtool @dbtool.bsky.social · 03/09/2026

We ran our sex-of-participants model over all of human PubMed: 12.9M studies, 1960–2025. The mono-sex gap has flipped. 1960s: 46% male-only vs 22% female-only. 2020s: female-only (18%) overtakes male-only (17%) for the first time. #metascience #bibliometrics

100

dbtool @dbtool.bsky.social · 03/09/2026

Mapping the sex & age of study participants across all of human PubMed (~13M abstracts) with a BiomedBERT model trained on MeSH check-tags (sex accuracy ~90%). Preliminary, on 7M studies so far: 56% both sexes, 23% male-only, 21% female-only. #metascience #bibliometrics

000

dbtool @dbtool.bsky.social · 01/09/2026

What we're building now: a model that reads a biomedical abstract and tells you WHO was studied — female only, male only, both, or not reported. Trained on 7M MeSH-labelled abstracts. Half of human studies don't state participants' sex. Free for research when it ships. #metascience

000

dbtool @dbtool.bsky.social · 01/09/2026

Where this is going: releasing these models openly, long-term. @nrobinsongarcia.bsky.social's call asks for transparent, comprehensive, freely accessible methods — we meet two of three. The API fees fund the datasets and models that get us toward the third. Research use is already free.

010

dbtool @dbtool.bsky.social · 01/09/2026

Right — the databases have no gender field. So studies either infer it from names, or hand-code a sample, with its own representativeness problems. Either way the error is rarely reported — and these numbers feed real policy debates, the leaky pipeline above all. That's where the stakes are.