# Benchmarks

Endpoints for exploring AI benchmark evaluation data. The benchmarks dataset contains 13,000+ questions from 9 benchmark suites, each mapped to occupational skill requirements to provide the empirical foundation for Demonstrated Capability scores.

**Access:** Enterprise only

## List Benchmarks

```http
GET /benchmarks
```

List AI benchmark evaluations with optional filtering.

### Parameters

| Parameter   | Type   | Required | Default | Description                              |
| ----------- | ------ | -------- | ------- | ---------------------------------------- |
| `skill`     | string | No       | -       | Filter by skill name (partial match)     |
| `benchmark` | string | No       | -       | Filter by benchmark suite (e.g., "mmlu") |
| `limit`     | int    | No       | 20      | Max results per page (1-100)             |
| `offset`    | int    | No       | 0       | Pagination offset                        |

### Response

```json
{
  "total": 13017,
  "limit": 20,
  "offset": 0,
  "benchmarks": [
    {
      "input_id": "0",
      "benchmark_name": "mmlu",
      "subject": "abstract_algebra",
      "task_description": "Find the degree for the given field extension Q(sqrt(2), sqrt(3), sqrt(18)) over Q.",
      "skills_tested_count": 3
    }
  ]
}
```

### Examples

```bash
# Get benchmarks testing Programming skill
curl -H "Authorization: Bearer $API_KEY" \
  "https://api.maidenlabs.tools/benchmarks?skill=Programming&limit=20"

# Get MMLU benchmarks
curl -H "Authorization: Bearer $API_KEY" \
  "https://api.maidenlabs.tools/benchmarks?benchmark=mmlu&limit=50"

# Paginate through results
curl -H "Authorization: Bearer $API_KEY" \
  "https://api.maidenlabs.tools/benchmarks?limit=100&offset=100"
```

***

## Get Benchmark Detail

```http
GET /benchmarks/{input_id}
```

Get detailed information about a specific benchmark evaluation.

### Path Parameters

| Parameter  | Type   | Description        |
| ---------- | ------ | ------------------ |
| `input_id` | string | Benchmark input ID |

### Response

```json
{
  "input_id": "12345",
  "benchmark_name": "humaneval",
  "subject": "code_generation",
  "task_description": "Write a function that takes a list of integers and returns the sum of all even numbers.",
  "input_text": "def sum_even(numbers: List[int]) -> int:\n    \"\"\"Return sum of even numbers in the list.\"\"\"\n",
  "skills_tested": [
    {
      "element_id": "2.B.3.e",
      "name": "Programming",
      "type": "Skills",
      "relevance_score": 0.95
    },
    {
      "element_id": "1.A.1.b.4",
      "name": "Deductive Reasoning",
      "type": "Abilities",
      "relevance_score": 0.72
    },
    {
      "element_id": "2.A.2.a",
      "name": "Critical Thinking",
      "type": "Skills",
      "relevance_score": 0.68
    }
  ]
}
```

### Example

```bash
curl -H "Authorization: Bearer $API_KEY" \
  "https://api.maidenlabs.tools/benchmarks/12345"
```

***

## Benchmark Sources

Cubit's benchmark data comes from the Stanford HELM (Holistic Evaluation of Language Models) project. The dataset includes 13,017 questions across 9 benchmark suites:

| Benchmark     | Category    | Questions | Description                               |
| ------------- | ----------- | --------- | ----------------------------------------- |
| wmt\_14       | Translation | 5,000     | Machine translation evaluation            |
| legalbench    | Legal       | 2,047     | Legal reasoning tasks                     |
| natural\_qa   | Knowledge   | 2,000     | Open-domain question answering            |
| med\_qa       | Medical     | 1,000     | Medical knowledge questions               |
| gsm           | Math        | 1,000     | Grade school math problems                |
| mmlu          | Knowledge   | 565       | Subject knowledge across multiple domains |
| commonsense   | Reasoning   | 498       | Commonsense reasoning                     |
| narrative\_qa | Reading     | 470       | Narrative comprehension                   |
| math          | Math        | 437       | Mathematical problem solving              |

***

## How Benchmarks Connect to Skills

Each benchmark is analyzed to identify which cognitive requirements it tests:

```
Benchmark: "Solve this calculus problem..."
    |
    |-- Mathematics (Skill) - relevance: 0.95
    |-- Mathematical Reasoning (Ability) - relevance: 0.90
    |-- Deductive Reasoning (Ability) - relevance: 0.75
    |-- Number Facility (Ability) - relevance: 0.60
```

These relevance scores flow through to compute AI Resilience Scores for each requirement.

***

## Use Cases

### Validate AI Capability Claims

```bash
# What benchmarks test "Complex Problem Solving"?
curl -H "Authorization: Bearer $API_KEY" \
  "https://api.maidenlabs.tools/benchmarks?skill=Complex+Problem+Solving&limit=50"
```

### Explore Benchmark Coverage

```bash
# How many benchmarks test each skill?
# (Aggregate results programmatically)
curl -H "Authorization: Bearer $API_KEY" \
  "https://api.maidenlabs.tools/skills?sort_by=benchmark_match_count"
```

### Understand Score Provenance

```python
# Get skill profile to see benchmark coverage
skill = client.get_skill("2.B.3.e")  # Programming
print(f"Tested by {skill['benchmark_coverage']['total_benchmarks']} benchmarks")
print(f"Top benchmarks: {skill['benchmark_coverage']['top_benchmarks']}")

# Examine specific benchmarks
benchmarks = client.list_benchmarks(skill="Programming", limit=10)
for b in benchmarks["benchmarks"]:
    print(f"[{b['benchmark_name']}] {b['task_description'][:60]}...")
```

***

## Data Provenance

| Aspect              | Details                                 |
| ------------------- | --------------------------------------- |
| Source              | Stanford HELM v1.13.0                   |
| Total Evaluations   | 1,233,645                               |
| Benchmark Questions | 13,017                                  |
| AI Models           | 95                                      |
| Benchmark Suites    | 9                                       |
| Skill Mappings      | Cubit-annotated                         |
| Update Frequency    | Quarterly                               |
| Methodology         | LLM-assisted annotation with validation |

***

## Limitations

{% hint style="info" %}
**Benchmark Coverage Varies**

Some requirements have extensive benchmark coverage (e.g., Mathematics: 15,000+ benchmarks) while others have sparse coverage (e.g., Gross Body Coordination: \~50 benchmarks).

Requirements with sparse coverage have higher uncertainty in their AI Resilience Scores. The `benchmark_match_count` field indicates evidence strength.
{% endhint %}

{% hint style="info" %}
**Language Model Focus**

Current benchmarks primarily evaluate language models, not robotics or embodied AI. Physical abilities may show artificially high AI performance due to cognitive proxies being tested rather than actual physical capability.
{% endhint %}


---

# Agent Instructions: Querying This Documentation

If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter:

```
GET https://docs.maidenlabs.tools/api-reference/benchmarks.md?ask=<question>
```

The question should be specific, self-contained, and written in natural language.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.