dataset-evaluation（数据集评估）

SKILL.md --- dataset_evaluation

Skill Name

dataset_evaluation

Description

Evaluate a miner submission by performing two evaluation steps:

1. Content Consistency Evaluation
2. Structured Data Quality Evaluation

The evaluator receives 5 cleaned data samples, the structured JSON, and the dataset schema, then computes a final score for the miner.

Input

{
  "cleaned_data_list": [
    "cleaned_text_1",
    "cleaned_text_2",
    "cleaned_text_3",
    "cleaned_text_4",
    "cleaned_text_5"
  ],
  "structured_data": {
    "field1": "value",
    "field2": "value"
  },
  "dataset_schema": {
    "fields": [
      {"name": "title", "type": "string", "required": true},
      {"name": "author", "type": "string", "required": false},
      {"name": "date", "type": "string", "required": false},
      {"name": "url", "type": "string", "required": true}
    ]
  }
}

Evaluation Procedure

Step 1 --- Content Consistency Evaluation (Weight 40%)

Goal: determine whether the 5 cleaned texts represent the same underlying content.

Method

1. Normalize text

• remove HTML
• lowercase
• remove excessive whitespace

1. Compute pairwise similarity across the 5 texts

Recommended metrics:

• cosine similarity (embedding based)
• OR Jaccard similarity

1. Compute the average similarity score.

Output

content_consistency_score (0-100)

Suggested mapping:

avg_similarity >= 0.9 → 100 0.8 – 0.9 → 80 – 100 0.6 – 0.8 → 60 – 80 0.4 – 0.6 → 40 – 60 < 0.4 → < 40

Step 2 --- Structured Data Quality Evaluation (Weight 60%)

Using the verified cleaned content, evaluate the structured JSON.

Compute four sub-scores.

2.1 Field Completeness (30%)

Evaluate whether all required fields exist.

Formula:

completeness_score = (# required fields present / total required fields) * 100

2.2 Value Accuracy (40%)

Evaluate whether each field value is consistent with the cleaned data.

Examples:

• title appears in cleaned text
• author name appears in text
• url matches source

Scoring guideline:

exact match → 100 partially correct → 60-80 inconsistent → <50

2.3 Type Correctness (15%)

Evaluate whether values match schema types.

Examples:

string number boolean array

Formula:

type_score = (# correct types / total fields) * 100

2.4 Information Sufficiency (15%)

Evaluate whether the structured data misses obvious information present in the cleaned text.

Example:

Cleaned text contains:

title author date

But structured JSON only includes:

title

Then deduct score.

Guideline:

complete extraction → 100 minor missing info → 70–90 major missing info → <60

Structuring Quality Score

structuring_quality_score = completeness_score * 0.30 + value_accuracy_score * 0.40 + type_score * 0.15 + information_sufficiency_score * 0.15

Range:

0 – 100

Step 3 --- Final Miner Score

miner_score = content_consistency_score * 0.4 + structuring_quality_score * 0.6

Range:

0 – 100

Output Format

The evaluator must return:

{
  "content_consistency_score": 92,
  "structuring_quality_score": 85,
  "miner_score": 88.2,
  "details": {
    "completeness_score": 90,
    "value_accuracy_score": 88,
    "type_score": 100,
    "information_sufficiency_score": 80
  }
}

Evaluator Rules

The evaluator must follow these principles:

1. Be deterministic and reproducible
2. Base judgments only on provided inputs
3. Avoid hallucination
4. Penalize missing or inconsistent data
5. Return scores strictly in the 0--100 range