Zum Hauptinhalt springen
tsecurity.de LIVE
Echtzeit-Radar & Feeds
Alle RSS Feeds
👥 Community & Social
Sichere ProgrammierungFliproom: a room changeover is a content problem(21.09.2026 um 04:10 Uhr)
Sichere ProgrammierungNova Adiutrix: My Second Agent Built My First Project's To-Do List(21.09.2026 um 04:11 Uhr)
IT Security ToolsAntiphishing v35456910988(21.09.2026 um 02:35 Uhr)
IT Security Toolsbrave-browser v1.98.12(21.09.2026 um 03:35 Uhr)
Sichere ProgrammierungFliproom: a room changeover is a content problem(21.09.2026 um 04:10 Uhr)
Sichere ProgrammierungNova Adiutrix: My Second Agent Built My First Project's To-Do List(21.09.2026 um 04:11 Uhr)
IT Security ToolsAntiphishing v35456910988(21.09.2026 um 02:35 Uhr)
IT Security Toolsbrave-browser v1.98.12(21.09.2026 um 03:35 Uhr)
Intelligence View
⚡ tsecurity.de Intelligence

🧞‍♂️Transform unstructured PDFs Job Offers into a dataset w. gemma4:2b

Reagiere als Erste:r — dein Feedback zählt!

This is a submission for the Gemma 4 Challenge: Build with Gemma 4

🤔 About the power of collections and our ability to compare things

First a bit of philosophy.

Did you notice how we tend to align things, tend to shape things so they can be aligned, compared (based on a common attributes like color, weight,...).

Comparing objects is much easier when they share common structure, then we can use attributes to get more knowledge, produce KPIs, make clever choices and put things in evidence.

👉 Well the same happens with machines : it's much much easier to compare and manage documents when they share a same structure.

This is the very core idea that motivated this work to explore how Open Source AI could help in a very pragmatic way... and feel the opportunities it opens with concrete prototypes.

🙋 What I Built

I've built a whole real life and live data pipeline that takes as input Open Data Public Sector Job offers (dataset/avis-de-vacances-de-poste-avp-drhfpnc) :

  • csv
  • Raw PDFs

Then,

  1. From a dedicated GH repo adriens/avps I've prepared a whole structured mix of md thanks to csv
  2. Then from the GH Action I did transform brut raw PDFs with pypi.org/marker-pdf into markdown
  3. I ended to publish a dedicated Zensical gh-pages website : adriens.github.io/avps

Next, this is where things go really interesting : I wanted to be able to compare job offers the one against the others... but markdown were far too much different the one from others :

  • Not the same number of sections
  • Not the same section titles : hard skills, soft skills, missions,...
  • Not the same levels of sections
  • Not necessarly itemized the one
  • Not the same style at all (section levels, CAPITALs, email, cities...)

... which made it very hard... or even impossible to compare them amongs the others... or even crazier : put them in a traditional SQL structured database.

👉 This is where gemma4:2b comes in to create a very well and consistent set of markdowns that can then be used for various use cases :

  • Create very well structured ePub to read job offers on the go (and docx)
  • Create a very clean and well organized PDF : very easy to load in assistants, print or to drop in any assistant
  • Deliver structured data with clean json files
  • Make a duckdb database and perform SQL on the data by using the now well structured markdowns, which made it possible to open unprecedented and exciting reporting opportunities (here in duckdb)
  • Share all this as a dataset on Kaggle
SELECT '--- RÉPARTITION DES COMPÉTENCES PAR DOMAINE ---'
as titre_report;
SELECT domaine, count(*) as nb_competences 
FROM savoir_faire 
GROUP BY 1 ORDER BY 2 DESC;

// Detect dark theme var iframe = document.getElementById('tweet-2055533578021396775-622'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2055533578021396775&theme=dark" }

🎯 Problems it solves

In input we really had very various kind of PDF documents, and no structured tabular data, now, they both are delivered as :

  • Well formated and structured markdown
  • A real database that embeds data as tables and views for advanced SQL reporting and charting
  • Ready to use and perfectly well structred ePub and PDF documents, very easy for LLMs to understand

🤗 Experience it creates

The experience is rather an data experience as thanks to data normalisation and standardization we can load and compare job offers, which make job search and indexing much much more efficient, whatever the input.

Last but not least, using gemma4:2b-it proves that great things can be achieved even with small resources and that well prepared data opens so many intelligence opportunities, without having to deal with frontier models as "the output I got is good enough".

🍿 Demo

💰 The benefits : then and now

Below the benchmark of markdown before and after

⚖️ Benchmark : marker-pdf vs. marker-pdfgemma4:e2b

Below some results:

Structure consistency:

📊 Analytics on top of database

One the well-structured json could be produced from the markdown I could efficiently load them into a duckdb database and do some reporting see AVPS DRHFPNC - Les pdf en SQL avec duckdb Kaggle notebook :

// Detect dark theme var iframe = document.getElementById('tweet-2055751934301380738-928'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2055751934301380738&theme=dark" }

📜 Code

🎁 Goodies

💡 How I Used Gemma 4

I chose google/gemma-4/transformers/gemma-4-e2b-it from kagglehub as I had a huge amounf of data to load (all New-Caledonia ones) and a restricted amount of time on Kaggle as well as small GPUs.

Also my intent was to be able to run this code one day onPrem on my very own hardware so I decided to stay as little as possible.

🤔 What remains to do...

Try to:

  • Add an evaluation phasis to check output consistency
  • Try to switch to CPU mode so the Notebook can be scheduled without exceeding the maximum Kaggle window
  • Use gemma-4-E4B and benchmark output quality
  • Produce native adoc with proper annotations
  • First produce json (and more standardized values, enums,...) then re-generate md/adoc from it
Ähnliche Beiträge
🔍 Verwandte News

Auch interessante Nachrichten 🧞‍♂️Transform unstructured PDFs Job Offers into a dataset w. gemma4:2b

Thematisch verwandte Begriffe: Transform, unstructured, PDFs, Offers · 6 Treffer

Laden...

Videos werden geladen ...

Laden...

Beiträge werden geladen ...

Laden...

Videos werden geladen ...

Laden...

Beiträge werden geladen ...

Laden...

Videos werden geladen ...

Laden...

Beiträge werden geladen ...

Laden...

Videos werden geladen ...

Laden...

Beiträge werden geladen ...

Laden...

Videos werden geladen ...

Zum Aktualisieren ziehen
ZERO-DAY CVE-2026-93977 | A vulnerability was determined in code-projects Assessment Management 1.…
Advisory →
TTS Reader • tsecurity.de Voice
tsecurity.de Icon
tsecurity.de App
Offline-Lesen, Eilmeldungen & 0ms Ladezeit

Installiere tsecurity.de direkt auf deinen Home-Bildschirm für das ultimative Vollbild-Magazinerlebnis ohne Browser-Leisten.

Nächster Beitrag
Themen-Radar & Intelligence Matrix
Echtzeit-Taxonomie nach Angriffsvektoren & Plattformen

tsecurity.de Live Threat Radar

🔴 LIVE RADAR
MONITORING
AKTIV
CVE-DATENBANK
LIVE
🔍
Community Radar & Live Chat
Sentinel Bot online • Live-Stream
Dein Cluster: Security Explorer
Match:
lädt…
Verbindung zum Community-Stream wird aufgebaut...
Bearbeitungsmodus — Senden überschreibt deine Nachricht
Community-Puls — was gerade passiert
lädt…
Aktivitäten deiner Analysten
lädt…
Neues Thema oder Eilmeldung einreichen

Reiche interessante Links, Zero-Days oder Debatten ein. Die Community entscheidet per Upvote über die Veröffentlichung.

Heiß diskutierte Einreichungen
🔖 Gespeicherte Artikel
📂 Keine gespeicherten Artikel vorhanden.
Zurück Ziehen Vor
Links: vorheriger Artikel Rechts: nächster Artikel unten: schließen
News NIS-2 Frühwarnung Tier-1 Intel ⏱️ 3 Min vor 10 Min
Artikeldaten werden geladen...

Zurück: vorheriger Vor: nächster
↗ Original-Quelle
Social Reaktionen Deine Reaktion zählt
Einstufung & Relevanz-Poll 0 Stimmen
In sozialen Netzwerken teilen 1-Klick