Skip to main content

Source: oceanid/@docs/guides/SME/workflow.md | ✏️ Edit on GitHub

SME Workflow Guide (v1 now, v2 planned)

Audience: Subject Matter Experts working with text, PDFs, images, and CSVs.

Goal: Show what to upload, how pre-labels appear, how to save, and where your work goes. CSVs are supported now (v1). Your existing pandas spreadsheet cleaners will be integrated next (v2) so you won’t run them manually.

Related guide (DB access + tables): SME Guide — Uploading, Annotating, and Accessing Raw Data (Staging)

What You Do​

  • Open Label Studio at label.<base> (Cloudflare Access).
  • Create/open a project and import your data (Text, PDFs, CSVs, images).
  • Review pre-labels (auto suggestions), correct spans/labels, Save.

Supported Inputs (v1)​

  • Text: gets NER pre-labels (entities like VESSEL, IMO, FLAG, PORT, DATE).
  • PDFs: Docling text extraction + NER pre-labels.
    • New: Draw boxes on pages (Rectangle tool) for TABLE/FIGURE/SECTION/SIGNATURE/STAMP. Boxes store to stage.pdf_boxes and are also converted to PDF-point coordinates for training.
  • CSVs: header-aware flattening + NER pre-labels (details below).
  • XLSX: header-aware flattening + NER pre-labels.
  • Images:
    • If images contain text, convert to PDF or OCR to text to get pre-labels.
    • Pure image annotations are supported in LS, but no auto pre-labels in v1.

How Pre‑Labels Work​

CSV/XLSX Instructions (v1)​

  • Prepare your CSV with at least one of:
    • text column: a sentence/paragraph to annotate for entities, or
    • pdf or url column: a public PDF link (the system fetches it for pre-labels)
  • Keep all other columns; they are preserved as metadata for provenance and later use.
  • Examples:
text,pdf,meta_id
"Vessel IMO 9395044 entered Singapore on 2024-09-01",,doc-123
,"https://example.com/report.pdf",doc-124
  • Import steps:
    • Project → Import → Upload CSV.
    • Mapping: map your text column to “text”; keep pdf/url columns (no special mapping required).
    • The ML backend is already wired to our adapter; open a task to see pre-labels.
  • Best practices:
    • If rows don’t have a single “text” column, create a temporary text column (in Excel/Sheets, no scripts) that combines the most useful fields (e.g., “Vessel {VESSEL_NAME} (IMO {IMO}) arrived {DATE} at {PORT}”).
    • For PDFs, ensure the pdf/url column contains a working HTTPS link.
    • For XLSX, the system reads all sheets/cells and generates header-aware “Header: Value” lines for NER.

Annotating and Saving​

  • Adjust spans and labels to be accurate and specific (e.g., prefer IMO over generic VESSEL where applicable).
  • Click Save/Submit. Your annotation is recorded immediately.

Draw Boxes on PDFs (Hybrid UI)​

  • Open a PDF task. You’ll see a “PDF Page (Boxes)” section with an image preview and the rectangle tool.
  • Use the labels TABLE, FIGURE, SECTION, SIGNATURE, or STAMP for regions of interest.
  • Save.

What happens under the hood:

  • Label Studio provides page images for the UI only; the system stores normalized box geometry and converts it to PDF-point coordinates using the original PDF size.
  • Your boxes land in stage.pdf_boxes and can be queried directly (latest rows in stage.v_pdf_boxes_latest).

Where Your Work Goes​

  • Hugging Face dataset (versioned JSONL): a permanent, auditable history of annotations used for model training.
  • Postgres stage schema:
    • stage.documents: task text and metadata (project/task IDs, source).
    • stage.extractions: the cleaned spans (label, value, offsets, confidence).
    • stage.table_ingest: raw table rows (CSV/XLSX) ingested on task creation via webhook.
    • stage.pdf_boxes: page regions you draw (both percent and PDF-point coordinates when available).
    • stage.v_pdf_boxes_latest: latest box rows for quick QA per page/label.
    • Views for QA/ops: stage.v_documents_freshness, stage.v_duplicates.

Label Set (What You’ll See)​

  • Entities aligned to our schema, e.g., VESSEL, VESSEL_NAME, IMO, MMSI, IRCS, FLAG, PORT, ORGANIZATION, PERSON, COMPANY, HS_CODE, SPECIES, DATE, LOCATION, RISK_LEVEL, etc.
  • These labels map downstream to curated tables so your work supports analytics and intelligence.

Quality Loop (Why Your Edits Matter)​

  • Your saved annotations update the HF dataset.
  • Nightly training exports a new model; the GPU server pulls it automatically.
  • Pre-labels improve over time, reducing your manual edits.

Images​

  • To get pre-labels from images containing text, convert to PDF or run OCR to text, then import that into LS.
  • For pure image annotation (boxes/regions), use LS image tools; pre-labels are not automated in v1.

Coming in v2: Your Pandas Cleaners, Integrated​

  • What changes for you:
    • You won’t run scripts locally. The system will run your existing pandas cleaners automatically in the cluster.
    • Cleaners will normalize per-country spreadsheets (encodings, delimiters, dates), compose a high-signal text column, and import clean tasks to LS on a schedule or on demand.
  • What stays the same:
    • You keep using LS to review pre-labels, correct spans, and save.
    • Your work still flows to the HF dataset and Postgres stage for QA and downstream use.

Do’s and Don’ts​

  • Do:
    • Provide a text or pdf/url column in CSVs to enable pre-labels.
    • Keep spans tight to the entity text; use the most specific label.
    • Add a simple “text” column in Excel/Sheets if needed; no scripts required in v1.
  • Don’t:
    • Upload sensitive content without clearance.
    • Introduce ad-hoc labels without checking the project schema.