One Schema for Every Eval

NickDouglas1 pts0 comments

Every Eval Ever | EvalEval Coalition

Infrastructure @EvalEval

One schema for<br>Every Eval Ever.

A unified, open data format and public dataset for AI evaluation results.<br>We are collecting all evaluation results in a standardized schema to enable rigorous research and broader impact.

Explore Schema

Launch Announcement

Developed by the EvalEval Coalition

with feedback from researchers at

Inspect

Why this Schema?

Addressing the fragmentation in AI evaluation to enable trust and comparability.

Unifying the Ecosystem

Evaluation results are currently siloed by framework. This schema creates a common interchange format, allowing results from HELM, EleutherAI, Inspect, and custom scripts to co-exist and be compared directly without complex mapping.

Trust through Provenance

A score without configuration is just noise. We go beyond the metric to capture the full experimental context—prompt templates, inference parameters, and system states—making every result traceable, transparent, and reproducible.

Actionable Meta-Science

Liberating evaluation results from static PDFs and closed leaderboards. We transform scattered metrics into a structured, queryable global dataset, powering the next generation of meta-analysis and automated leaderboard construction.

The Schema

A granular, line-by-line breakdown of the standardized format.

Aggregate Schema<br>Instance Schema

59ee0934-f60d-4d4b-b986-844fc51e89a3.json

"schema_version": "0.2.2",<br>"evaluation_id": "lm-eval/moonshotai.../17642...",<br>"retrieved_timestamp": "1764204739.50717",

"source_metadata": {<br>"source_name": "lm-eval",<br>"source_type": "evaluation_run",<br>"source_organization_name": "EleutherAI",<br>"source_organization_url": "https://www.eleuther.ai",<br>"evaluator_relationship": "third_party"<br>},

"model_info": {<br>"name": "Kimi K2 Instruct",<br>"id": "moonshotai/kimi-k2-instruct",<br>"developer": "moonshotai",<br>"inference_platform": "Together AI",<br>"inference_engine": {<br>"name": "vLLM",<br>"version": "0.6.3"<br>},

"eval_library": {<br>"name": "lm-eval",<br>"version": "0.4.11"<br>},

"evaluation_results": [<br>"evaluation_result_id": "mmlu_pro_cot_correct",<br>"evaluation_name": "MMLU-Pro - COT correct",<br>"evaluation_timestamp": "1764204739.50717",<br>"source_data": {<br>"source_type": "hf_dataset",<br>"dataset_name": "MMLU-Pro",<br>"hf_repo": "TIGER-Lab/MMLU-Pro",<br>"hf_split": "test",<br>"samples_number": 12032<br>},<br>"metric_config": {<br>"evaluation_description": "Fraction of correct answers after chain of thought",<br>"lower_is_better": false,<br>"score_type": "continuous",<br>"min_score": 0.0,<br>"max_score": 1.0<br>},<br>"score_details": {<br>"score": 0.819,<br>"details": {<br>"tab": "Accuracy"<br>},<br>"generation_config": {<br>"generation_args": {<br>"temperature": 0.0,<br>"top_p": 1.0,<br>"top_k": -1,<br>"max_tokens": 2048,<br>"reasoning": true<br>},<br>"additional_details": {<br>"description": "Chain of thought prompting used."<br>],

"detailed_evaluation_results": {<br>"file_path": "detailed_results/e70acf51....jsonl",<br>"format": "jsonl",<br>"checksum": "a1b2c3d4..."

Metadata & ID

Every file starts with versioning and a unique identifier.

schema_version: Now v0.2.2.

evaluation_id: Unique ID.

retrieved_timestamp: Unix Time.

Source Provenance

Links to original data and distinguishes source types.

Model Specification

Complete model specification using Hugging Face format.

name: Human-readable model name.

id: HuggingFace model ID.

inference_platform: API provider.

inference_engine: Local runner.

Evaluation Library

Mandatory field specifying the framework used, like HELM, lm-eval, or custom code.

Flexible Metrics

Supports continuous scores and level-based ratings.

evaluation_result_id: Stable key linking to instance records.

metric_config: Configuration for the metric.

score_details: The actual score value.

Generation Config

Reproducibility parameters for the generation.

Detailed Results

This links to the separate Instance Level file.

detailed_evaluation_results: Object with file path.

checksum: Verification hash.

"schema_version": "instance_level_eval_0.2.2",<br>"evaluation_id": "lm-eval/moonshotai...",<br>"model_id": "moonshotai/kimi-k2-instruct",<br>"evaluation_name": "mmlu_pro",<br>"sample_id": "test_1042",<br>"interaction_type": "single_turn",

"input": {<br>"raw": "Question: Which is a scalar?...",<br>"reference": ["C"],<br>"choices": ["Velocity", "Force", "Speed"]<br>},

"output": {<br>"raw": ["Speed is a scalar. The answer is C."],<br>"reasoning_trace": ["Velocity has direction..."]<br>},

"answer_attribution": [<br>"turn_idx": 0,<br>"source": "output.raw",<br>"extracted_value": "C",<br>"extraction_method": "regex",<br>"is_terminal": true<br>],

"evaluation": {<br>"score": 1.0,<br>"is_correct": true<br>},<br>"token_usage": {<br>"input_tokens": 365,<br>"output_tokens": 42,<br>"total_tokens": 407

Instance Meta

Identifier for this specific sample and interaction type.

interaction_type: single_turn, multi_turn, or agentic.

Input Data

The prompt and ground truth info.

Output

The models raw response and optional reasoning trace.

Evaluation & Usage

Instance level score, correctness and token stats.

Design...

schema eval evaluation results format score

Related Articles