Skip to content

Collections

Upload Variables

Elluminate supports batch-adding variables to a Template Variables Collection from a file. This is the fastest way to create or expand test sets spanning happy paths, edge cases, and adversarial inputs.

Supported formats - CSV: Header row required. UTF‑8, comma delimited. - JSONL: One JSON object per line. - Excel: .xlsx or .xls; first row contains headers.

Header rules - Columns become the collection’s variable names. Use concise, stable names like customer_name, plan_tier. - You can introduce new columns during upload; the collection’s schema updates automatically. - Missing cells become empty strings "". All values are stored as strings. - Files must contain at least one data row; only a header is rejected.

Duplicates - Duplicate rows within the same upload are automatically deduplicated. - Exact duplicates already present in the collection are allowed; handle downstream (e.g., sorting or filtering) if you want uniqueness across time.

Read‑only collections - Uploads are blocked for locked collections. Unlock to modify and re‑lock to preserve comparability across experiments.

CSV example 

name,age,occupation
John Doe,30,Engineer
Jane Smith,28,Designer

JSONL example 

{"name": "John Doe", "age": "30", "occupation": "Engineer"}
{"name": "Jane Smith", "age": "28", "occupation": "Designer"}

Excel example - Sheet1, row 1 is the header: name | age | occupation - Rows 2..N contain values.

Validation and parsing - CSV: comma delimiter; invalid structure (e.g., ragged rows) is rejected. - JSONL: each line must be valid JSON; parsing errors are rejected. - Excel: first sheet parsed; invalid content causes a parse error.

Tips for evaluation design - Use diverse distributions across columns (length, language, edge tokens) to stress specific behaviors. - Encode adversarial triggers as explicit columns (attack_type, red_team_scenario) to analyze by slice. - Keep column names stable; add new columns instead of overloading existing ones.

Export Variables

You can export a collection’s variables for external analysis or reuse.

Formats - JSONL: application/jsonlines; one object per line. Good for programmatic pipelines. - CSV: text/csv; human‑readable; preserves column order. - XLSX: Excel workbook; best for spreadsheet workflows.

Filenames - Downloads include the collection name and format, e.g., MyCollection_variables.csv.

Troubleshooting

  • File contains no data: Ensure you have at least one row below the header.
  • Invalid CSV structure: Keep a consistent number of columns per row; use comma delimiter.
  • Invalid JSONL: Each line must be a standalone JSON object; avoid arrays or trailing commas.
  • Read‑only error: Unlock the collection to allow changes, then re‑lock.
  • Unexpected empty values: Empty cells are normalized to empty strings; treat null‑ish values explicitly if you need different semantics.

API Endpoints

  • Upload: POST /api/v0/projects/{projectId}/collections/{collectionId}/entries/batches
  • Export: GET /api/v0/projects/{projectId}/collections/{collectionId}/files?format={jsonl|csv|xlsx}