DATA FOUNDRY · AI BENCHMARKS
Empirical foundation model benchmark rows and catalog metadata.
Structured evaluations from the Open LLM Leaderboard combined with verified provider catalog parameters across 460+ production models in clean Apache Parquet.
COMMERCIAL ACCESS
Verified Parquet archives + Claude MCP integration
- ✓ Verified Data Foundry Parquet partitions (zstd)
- ✓ Model Context Protocol (MCP) tool integration
- ✓ Metering: Inquire for Access
- ✓ Deterministic quality gate & hash verification
- ✓ Contractual commercial rights clearance
Quotes reviewed by systems engineering within 1 business day
COLUMN SPECIFICATIONS
Typed relational schema.
Columnar Parquet & DuckDB ready.
| Column Name | Data Type | Description |
|---|---|---|
model_slug |
VARCHAR(128) | Canonical model repository identifier |
eval_metric |
VARCHAR(64) | Evaluated benchmark task (e.g. MMLU, GSM8K, MATH, IFEval) |
normalized_score |
FLOAT | Normalized evaluation score (0.0 to 100.0) |
context_window |
INT32 | Maximum supported context token length |
pricing_prompt_usd |
DOUBLE | Reported input cost per million tokens |
pricing_completion_usd |
DOUBLE | Reported output cost per million tokens |
SYSTEM STATUS & CAPABILITY BOUNDARIES
Engineering reality.
Verified operational features vs planned protocol capabilities.
Verified Operational Today
- ✓ Apache Parquet Partitions: Columnar archives (Zstandard compressed) with typed relational schemas.
- ✓ Zero-Copy Local Analytics: In-memory SQL exploration via DuckDB, Python Polars, and Arrow.
- ✓ Cryptographic Deduplication: SHA-256 release manifests and seen-keys ledger preventing record drift.
- ✓ Statutory Rights Clearance: 100% license-cleared provenance with statutory citations.
- ✓ Technical Quote Intake: Systems engineering review and schema sample verification.
Pilot & In-Specification Capabilities
- ◌ Remote MCP Streaming Adapter:
foundry_get_model_benchmarktool endpoint available under enterprise pilot agreements. - ◌ Automated Micro-Settlement: Agent-to-agent procurement and escrow via EXCHANGER (active protocol specification).
- ◌ Multi-Region Edge Replicas: Global Cloudflare R2 bucket replication currently in private validation.
VERIFIED SAMPLE RECORD
Inspect the raw payload.
Zero mocked fields.
{
"model_slug": "meta-llama/Llama-3.3-70B-Instruct",
"eval_metric": "MMLU-Pro",
"score": 68.4,
"context_window": 131072,
"input_cost_m_tokens": 0.40,
"output_cost_m_tokens": 0.80,
"architecture": "transformer_dense"
}
AGENT INTERFACE & MCP TOOL CONTRACT
Plug directly into Claude Code, Codex, and autonomous workflows.
Deterministic data retrieval.
Machine-Readable Contract
Direct dataset extraction is delivered via high-throughput columnar Apache Parquet. For autonomous assistants, ELYRAON provides standard Model Context Protocol (MCP) tool schemas and REST integration adapters with signed SHA-256 integrity receipts.
Adapter identifier: foundry_get_model_benchmark
Pricing status: Inquire for Access
Active MCP Endpoint: https://elyraon.com.co/agents/mcp (Host to live structural verification tools; metered dataset adapters available in scoped pilots).
{"tool": "foundry_get_model_benchmark", "arguments": {"model_id": "meta-llama/Llama-3.3-70B-Instruct"}}
MACHINE-READABLE SPECIFICATIONS
Structured metadata for autonomous agents.
Zero scrapers, zero hallucinations.
{
"entity_id": "https://elyraon.com.co/datasets/ai-benchmarks-evals/#dataset",
"catalog_id": "https://elyraon.com.co/datasets/#catalog",
"canonical_url": "https://elyraon.com.co/datasets/ai-benchmarks-evals/",
"primary_format": "application/vnd.apache.parquet",
"compression": "zstd",
"ingestion_pipeline": "ai_benchmarks.py",
"mcp_tool_name": "foundry_get_model_benchmark",
"license": "Creative Commons Attribution 4.0 (CC BY 4.0) & Public Terms",
"duckdb_quickstart": "SELECT * FROM read_parquet('ai-benchmarks-evals.parquet') LIMIT 5;"
}
WHY ELYRAON DATA FOUNDRY
Architectural contrast.
Built for engineering reality.
Combines academic evaluation scores with live market deployment pricing to help engineering teams optimize model router economics.
Conventional Data Feeds
- Undetected socket drops during volatility
- Messy unstructured CSV/JSON with broken parsers
- Ambiguous copyright terms creating lawsuit liability
- Expensive multi-thousand dollar rigid contracts
ELYRAON Data Foundry
- Continuous quality gate reconciliation
- Clean, typed, content-addressed Apache Parquet
- 100% license-cleared provenance with statutory citations
- Clear request-quote pricing and cryptographic delivery verification
RELATED DATASETS & AGENT PLATFORM
Connected intelligence feeds.
Cross-domain discovery and autonomous infrastructure.
TECHNICAL FREQUENTLY ASKED QUESTIONS
Clear answers.
Engineering boundaries.
What is the verified dataset inventory?
1,319 benchmark evaluation rows and 464 production model catalog records.
Can this data be used to train model routing agents?
Yes. Many teams use this dataset to train cost-vs-accuracy arbitration models.
Are evaluations hardware-pinned?
Benchmark scores reflect standardized community leaderboard submissions with documented execution environments.
What formats are provided?
Apache Parquet lake partitions and on-demand JSON query endpoints.
How do I request access?
Submit an inquiry with your organization and intended use case.