golden-dataset

# Golden Dataset

Comprehensive patterns for building, managing, and validating golden datasets for AI/ML evaluation. Each category has individual rule files in `rules/` loaded on-demand.

## Quick Reference

| Category | Rules | Impact | When to Use |
| -------- | ----- | ------ | ----------- |
| [Curation](#curation) | 2 | HIGH | Content collection, annotation pipelines |
| [Management](#management) | 2 | HIGH | Versioning, backup/restore |
| [Validation](#validation) | 1 | CRITICAL | Regression testing |
| [Add Workflow](#add-workflow) | 1 | HIGH | 9-phase curation, quality scoring, bias detection, silver-to-gold |

Total: 6 rules across 4 categories. House thresholds and scars: `references/ork-delta.md`.

## Curation

Content collection, multi-agent annotation, and diversity analysis for golden datasets.

| Rule | File | Key Pattern |
| ---- | ---- | ----------- |
| Collection | `rules/curation-collection.md` | Content type classification, quality thresholds, duplicate prevention |
| Annotation | `rules/curation-annotation.md` | Multi-agent pipeline, consensus aggregation, Langfuse tracing |

Difficulty ladder, coverage floors, and duplicate thresholds: `references/ork-delta.md`.

## Management

Versioning, storage, and CI/CD automation for golden datasets.

| Rule | File | Key Pattern |
| ---- | ---- | ----------- |
| Versioning | `rules/management-versioning.md` | JSON backup format, embedding regeneration, disaster recovery |
| Storage | `rules/management-storage.md` | Backup strategies, URL contract, data integrity checks |

CI automation for backups is upstream's job; see "Upstream coverage" below.

## Validation

Quality scoring, drift detection, and regression testing for golden datasets.

| Rule | File | Key Pattern |
| ---- | ---- | ----------- |
| Regression | `rules/validation-regression.md` | Difficulty distribution, pre-commit hooks, full dataset validation |

Schema validation and duplicate detection are upstream's job (see "Upstream coverage"
below); the house thresholds they must enforce live in `references/ork-delta.md`.

## Add Workflow

Structured workflow for adding new documents to the golden dataset.

| Rule | File | Key Pattern |
| ---- | ---- | ----------- |
| Add Document | `rules/curation-add-workflow.md` | 9-phase curation, parallel quality analysis, bias detection |

## Quick Start Example

```python
async def validate_before_add(document: dict, source_url_map: dict) -> dict:
    """Pre-addition validation for golden dataset entries."""
    errors = []

    # 1. URL contract check
    if "placeholder" in document.get("source_url", ""):
        errors.append("URL must be canonical, not a placeholder")

    # 2. Content quality
    if len(document.get("title", "")) < 10:
        errors.append("Title too short (min 10 chars)")

    # 3. Tag requirements
    if len(document.get("tags", [])) < 2:
        errors.append("At least 2 domain tags required")

    return {"valid": len(errors) == 0, "errors": errors}
```

## Key Decisions

| Decision | Recommendation |
| -------- | -------------- |
| Backup format | JSON (version controlled, portable) |
| Embedding storage | Exclude from backup (regenerate on restore) |
| Quality threshold | >= 0.70 quality score for inclusion |
| Confidence threshold | >= 0.65 for auto-include |
| Duplicate threshold | >= 0.90 similarity blocks, >= 0.85 warns |
| Min tags per entry | 2 domain tags |
| Min test queries | 3 per document |
| Difficulty balance | Trivial 3, Easy 3, Medium 5, Hard 3 minimum |
| CI frequency | Weekly automated backup (Sunday 2am UTC) |

## Common Mistakes

1. Using placeholder URLs instead of canonical source URLs
2. Skipping embedding regeneration after restore
3. Not validating referential integrity between documents and queries
4. Over-indexing on articles (neglecting tutorials, research papers)
5. Missing difficulty distribution balance in test queries
6. Not running verification after backup/restore operations
7. Testing restore procedures in production instead of staging
8. Committing SQL dumps instead of JSON (not version-control friendly)

## Running a dataset as an experiment

Curating a dataset is half the job; the other half is running something against it and scoring the
result. Both Langfuse SDKs ship a runner, and their shapes differ.

**Python (SDK 4.x):** see `monitoring-observability/references/experiments-api.md`.

**JS/TS (SDK 5.x):** `@langfuse/client` exposes the runner directly on a fetched dataset.

```typescript
import { LangfuseClient } from "@langfuse/client";

const langfuse = new LangfuseClient();
const dataset = await langfuse.dataset.get("my-evaluation-dataset");

const result = await dataset.runExperiment({
  name: "Retrieval quality",
  task: myTask,               // (params) => Promise<any>
  evaluators: [myEvaluator],  // per-item: (params) => Promise<Evaluation | Evaluation[]>
});
```

| Type | Scores | Use for |
|---|---|---|
| `Evaluator` | one item | Per-example quality (faithfulness, relevance) |
| `RunEvaluator` | the whole run | Aggregate assertions — pass rate, mean score, regression checks |
| `Evaluation` | — | `{ name, value, comment?, metadata?, dataType?, configId? }` |

A per-item `Evaluator` cannot see the other items, so anything comparative belongs in a
`RunEvaluator`. `createEvaluatorFromAutoevals` wraps an autoevals scorer instead of hand-writing
one, and `RegressionError` is thrown when a run regresses against a configured baseline — catch it
to fail CI on a quality drop rather than only on an exception.

Full JS surface: `monitoring-observability/references/langfuse-js-v5.md`.

## Evaluations

See `test-cases.json` for 9 test cases across all categories.

## Upstream coverage (do not restate)

| Topic | First-party source |
| ----- | ------------------ |
| Dataset schema validation (JSON Schema, field constraints) | https://json-schema.org and https://zod.dev |
| Duplicate detection via embeddings, cosine similarity | https://github.com/pgvector/pgvect