id stringlengths 6 15 | category stringclasses 4
values | difficulty stringclasses 3
values | prompt stringlengths 50 129 | expected unknown | scoring_method stringclasses 7
values | source stringclasses 1
value | version stringclasses 1
value |
|---|---|---|---|---|---|---|---|
json-001 | structured_output | easy | Return JSON with name set to Ada and active set to true. | {
"name": "Ada",
"active": true
} | json_schema | manual | 1.0 |
json-002 | structured_output | easy | Return JSON containing language as Python and version as 3.12. | {
"language": "Python",
"version": "3.12"
} | json_schema | manual | 1.0 |
json-003 | structured_output | medium | Return JSON with service set to api, replicas set to 3, and enabled set to true. | {
"service": "api",
"replicas": 3,
"enabled": true
} | json_schema | manual | 1.0 |
json-004 | structured_output | medium | Return JSON with a user object containing name as Sam and role as developer. | {
"user": {
"name": "Sam",
"role": "developer"
}
} | json_schema | manual | 1.0 |
json-005 | structured_output | hard | Return JSON with service as billing, regions as an array containing east and west, and monitoring as an object with enabled true. | {
"service": "billing",
"regions": [
"east",
"west"
],
"monitoring": {
"enabled": true
}
} | json_schema | manual | 1.0 |
json-006 | structured_output | hard | Return JSON with status as success and a result object containing count 2 and valid true. | {
"status": "success",
"result": {
"count": 2,
"valid": true
}
} | json_schema | manual | 1.0 |
text-match-001 | software_qa | easy | Explain dependency injection in one short sentence. | "Dependency injection provides dependencies to a component from the outside." | normalized_exact_match | manual | 1.0 |
qa-002 | software_qa | easy | What is dependency injection? Include the terms dependency injection and reusable dependencies. | {
"required_keywords": [
"dependency injection",
"reusable dependencies"
]
} | keyword_match | manual | 1.0 |
qa-003 | software_qa | medium | Explain why dependency injection helps unit testing. | "Dependency injection improves unit testing by allowing dependencies to be replaced with test doubles." | reference_text | manual | 1.0 |
qa-004 | software_qa | medium | Explain the benefit of separating business logic from infrastructure dependencies. | "Separating business logic from infrastructure dependencies improves maintainability and makes unit testing easier." | semantic_similarity | manual | 1.0 |
qa-005 | software_qa | hard | Explain dependency injection and mention loose coupling, testability, and dependency inversion. | {
"required_keywords": [
"dependency injection",
"loose coupling",
"testability",
"dependency inversion"
]
} | keyword_match | manual | 1.0 |
qa-006 | software_qa | hard | Explain how dependency injection supports maintainable applications. | "Dependency injection supports maintainability by separating object creation from business logic and making dependencies replaceable." | normalized_exact_match | manual | 1.0 |
code-001 | code_understanding | easy | Explain what Python type hints provide. Include type hints and static analysis. | {
"required_keywords": [
"type hints",
"static analysis"
]
} | keyword_match | manual | 1.0 |
code-002 | code_understanding | easy | Explain the purpose of a unit test in one sentence. | "A unit test verifies the behavior of a small, isolated unit of code." | normalized_exact_match | manual | 1.0 |
code-003 | code_understanding | medium | Explain why a function should validate its inputs. | "Input validation helps prevent invalid data from entering a function and makes failures easier to diagnose." | reference_text | manual | 1.0 |
code-004 | code_understanding | medium | Explain how dependency injection improves the testability of a service. | "Dependency injection allows a service to receive replaceable dependencies, making it easier to isolate and test its behavior." | semantic_similarity | manual | 1.0 |
code-005 | code_understanding | hard | Explain the benefits of static analysis and mention type errors, code quality, and early defect detection. | {
"required_keywords": [
"type errors",
"code quality",
"early defect detection"
]
} | keyword_match | manual | 1.0 |
code-006 | code_understanding | hard | Summarize the purpose of a Python code review in a short response. | {
"min_words": 5,
"max_words": 40,
"max_characters": 300,
"require_non_empty": true
} | output_constraints | manual | 1.0 |
instruction-001 | instruction_following | easy | Explain input validation in one sentence and include the phrases input validation and invalid data. | {
"required_phrases": [
"input validation",
"invalid data"
]
} | contains_all | manual | 1.0 |
instruction-002 | instruction_following | easy | Explain unit testing in a response of 5 to 25 words. | {
"min_words": 5,
"max_words": 25,
"max_characters": 200,
"require_non_empty": true
} | output_constraints | manual | 1.0 |
instruction-003 | instruction_following | medium | Explain error handling and include the phrases clear error messages and exception handling. | {
"required_phrases": [
"clear error messages",
"exception handling"
]
} | contains_all | manual | 1.0 |
instruction-004 | instruction_following | medium | Explain why API input validation matters in one sentence. | "API input validation helps reject invalid requests before they reach application logic." | normalized_exact_match | manual | 1.0 |
instruction-005 | instruction_following | hard | Explain reliable API design and include input validation, clear error messages, and consistent response formats. | {
"required_phrases": [
"input validation",
"clear error messages",
"consistent response formats"
]
} | contains_all | manual | 1.0 |
instruction-006 | instruction_following | hard | Explain how clear error messages help users and developers diagnose failures. | "Clear error messages describe what went wrong and help users and developers identify the cause of a failure." | reference_text | manual | 1.0 |
LLM Evaluation Benchmark
Overview
This dataset provides structured examples for evaluating language model outputs across multiple task categories and scoring methods.
It is intended for AI engineers, researchers, and developers building reproducible LLM evaluation pipelines.
Dataset contents
The dataset uses JSON Lines (JSONL). Each record represents one evaluation example.
Each example contains fields such as:
id: unique example identifiercategory: task categorydifficulty: difficulty levelprompt: input provided to the modelexpected: reference answer or scoring configurationscoring_method: evaluation methodsource: example provenanceversion: example version
Task categories
software_qastructured_outputinstruction_followingcode_understanding
Scoring methods
The examples can use scoring methods including:
- JSON schema validation
- Keyword matching
- Required phrase matching
- Normalized exact match
- Output constraints
- Reference-based text evaluation
- Semantic similarity
- LLM-as-judge evaluation
Some scoring methods require optional dependencies or additional configuration in the evaluation runner.
Intended uses
- Experimenting with LLM evaluation pipelines
- Testing structured output and instruction following
- Comparing model behavior on a consistent dataset
- Learning about evaluation metrics and benchmark design
Limitations
The dataset's coverage is limited by its size and example selection. Individual scoring methods measure different properties and do not independently establish factual correctness or overall model quality.
Scores should be interpreted alongside example-level results, evaluation configuration, model settings, and known metric limitations.
Demo-backend scores from the associated benchmark are for pipeline testing and must not be presented as real model performance.
Data validation
Validate the dataset using the associated benchmark project's schema and tests. Review examples for unique IDs, correct scoring configuration, clear prompts, and suitable expected outputs.
Provenance and licensing
Document the source and usage rights of all examples before redistributing the dataset. Add an explicit license only after confirming that you have the rights to release the dataset under that license.
Related project
Link the associated GitHub repository here after its public URL is available.
- Downloads last month
- 32