Showcase Cases Index

Juicer showcase cases, grouped by CDR-Bench-aligned capability. Each case uses a structured recipe to demonstrate a specific data-refinement operator or pipeline.

Category Overview

Category	File	Count	Capabilities
Atomic Mapper	atomic_mapper.jsonl	6	Single-step text transforms: copyright removal, whitespace normalization, email/IP cleaning, plus multi-step mapper chains
Atomic Filter	atomic_filter.jsonl	5	Single-step filter decisions: text length, alphanumeric ratio, character repetition, max line length
PII Redaction	pii.jsonl	10	Name / ID / address / contact / temporal placeholder redaction
Hallucination	hallucination.jsonl	5	Reference-grounded detection and correction (JSON output)
Rubric Scoring	rubric.jsonl	4	HelpSteer2 dimension scoring (complexity, verbosity, coherence, correctness)
Safety	safety.jsonl	6	Aegis safety-label classification (prompt_label, response_label) + multi-step combos
Order-sensitive	order_sensitive.jsonl	8	Multi-operator chains where execution order matters
Compositional	compositional.jsonl	7	Multi-step end-to-end pipelines (PII redaction combos, safety classification chains)

Total: 51 cases across 8 capability dimensions.

Design Principles

CDR-Bench-aligned operator taxonomy: cases follow CDR-Bench’s atomic_m / atomic_f / semantic_* / order_* track structure; each case maps to a clear operator and operator_kind.

Atomic-first, compositional as supplement: most cases are single-step atomic operations (atomic mapper / atomic filter), while compositional cases show multi-step operator coordination.

Structured prompt format: explicit numbered steps make operation order and output requirements easy to inspect.

Showcase-oriented: cases demonstrate the breadth of Juicer’s data-refinement capabilities across all major CDR-Bench operator categories.

Case Data Format (unified schema)

Each case is a JSONL line; all categories share the same schema:

{
  "id": "am-001",
  "capability": "atomic_mapper",
  "operator": "whitespace_normalization_mapper",
  "operator_kind": "mapper",
  "title": "Whitespace normalization",
  "description": "Atomic mapper: collapse multiple spaces, trim leading/trailing whitespace",
  "input_text": "  The   quick  brown   fox.  ",
  "reference": "(optional; for hallucination detection and other reference-grounded tasks)",
  "recipe": "Normalize whitespace: collapse multiple spaces to one, trim.",
  "output_format": "tagged_text",
  "expected_status": "KEEP",
  "expected_text": "The quick brown fox.",
  "why_good": "What capability this case demonstrates"
}

Field notes:

• output_format: tagged_text (outputs <status>KEEP|DROP</status><clean_text>...</clean_text>) or json (outputs structured JSON)

• operator_kind: mapper / filter / semantic_mapper / compositional

Per-Category Notes

Atomic Mapper (atomic_mapper.jsonl)

Single-step text-transform operators, plus short mapper chains for demonstrating CDR-Bench-style data transformations.

Cases:

• am-001 copyright removal, am-002 whitespace normalization, am-005 email cleaning

• am-003 / am-004 4-step pipelines (email + IP + links + more), am-006 2-step pipeline (IP + MAC)

Atomic Filter (atomic_filter.jsonl)

Single-step filter-decision operators that demonstrate threshold-based KEEP/DROP decisions.

Cases:

• af-001 / af-004 text length filter, af-002 alphanumeric ratio filter, af-003 character repetition filter, af-006 max line length filter

PII Redaction (pii.jsonl)

A natural advantage of local deployment: redact sensitive data without it leaving your machines.

Cases:

• pii-001 / pii-006 person names, pii-002 / pii-007 ID numbers, pii-003 / pii-004 addresses

• pii-005 contact info (email/phone), pii-008 / pii-009 / pii-010 temporal/birth-date info

Hallucination (hallucination.jsonl)

Reference-grounded comparison tasks, JSON output.

Cases:

• hall-001 / hall-003 / hall-006 hallucination correction (fix hallucinated content using reference)

• hall-004 / hall-007 hallucination detection (determine if text contains hallucinated content)

Rubric Scoring (rubric.jsonl)

HelpSteer2 dimension scoring: one atomic case per dimension.

Cases:

• rub-001 complexity, rub-002 verbosity, rub-003 coherence, rub-004 correctness

Safety (safety.jsonl)

Aegis safety-label classification: atomic per-field cases plus multi-step combos.

Cases:

• safe-001 / safe-005 prompt_label classification, safe-002 / safe-003 response_label classification

• safe-004 / safe-006 3-step pipelines (prompt_label + response_label + violated_categories)

Order-sensitive (order_sensitive.jsonl)

Multi-operator chains where the specified execution order determines the result. Showcases Juicer’s ability to follow recipe order precisely.

Cases:

• os-001 through os-008: diverse 3-5 step pipelines mixing filters and mappers (length filters, copyright cleaning, IP/MAC/links removal, word filters, punctuation normalization)

Compositional (compositional.jsonl)

Multi-step end-to-end pipelines, showcasing complex recipe handling.

Cases:

• comp-001 / comp-002 / comp-004 PII redaction combos (person + temporal, contact + location)

• comp-003 / comp-005 / comp-006 / comp-007 safety classification chains (prompt_label + response_label + violated_categories)