schema stringclasses 1
value | benchmark_suite_id stringclasses 1
value | canonical_board_url stringclasses 1
value | live_api_url stringclasses 1
value | verify_url stringclasses 1
value | did_signer stringclasses 1
value | doi stringclasses 1
value | statistical_separation_rule stringclasses 1
value |
|---|---|---|---|---|---|---|---|
every-eval-ever/submission/1.0 | csoai/gspc-33 | https://councilof.ai/benchmarks | https://councilof.ai/api/gspc | https://councilof.ai/verify | did:web:csoai.org#card-attestation-1 | 10.57967/hf/gspc-boards | Exact two-sided McNemar p < 0.05 on discordant items vs best third-party base model; Track A (Third-Party Base) vs Track B (CSOAI/MEOK Overlays) strictly separated. |
CSOAI GSPC-33 Living Benchmark Suite (csoai-gspc)
Council of AI (CSOAI Ltd, UK Companies House 16939677) — Independent AI-behaviour measurement body.
Live Scoreboard: https://councilof.ai/benchmarks · Live API: https://councilof.ai/api/gspc · Free Ed25519 Verifier: https://councilof.ai/verify
Doctrine: Measurement, never certification.
UNMEASUREDis first-class (33declared axes ·23measured on live board +10frozenn=32 >= 30subject-family candidate banks included in this repository).SEPARATEDrequires exact paired McNemar $p < 0.05$ on discordant items(b, c)against the best base model; otherwiseTIE.
1-Line Runnable Harnesses (Run GSPC on Your Model Today)
A. UK AISI Inspect AI (inspect_ai)
pip install inspect-ai "git+https://huggingface.co/datasets/csoai/gspc-benchmark-suite.git#subdirectory=pypi"
inspect eval csoai_gspc/governance --model openai/gpt-4o
B. EleutherAI lm-evaluation-harness (lm_eval)
git clone https://huggingface.co/datasets/csoai/gspc-benchmark-suite
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.3-70B-Instruct --include_path ./gspc-benchmark-suite/lm_eval/tasks/gspc --tasks gspc_governance
C. Kaggle Community Benchmarks (@kbench.task)
Import kaggle/kaggle_benchmark_notebook.py directly into Kaggle Benchmarks to evaluate frontier models on GSPC with paired McNemar separation and Ed25519 receipt verification.
D. Linux Foundation / AAIF OpenEnv (0.2.1)
Use openenv/openenv-manifest.json (reset(), step(), state()) for containerized multi-turn tool and sandbox boundary evaluation.
Measurement, not certification. Anyone can re-check at https://councilof.ai/verify
- Downloads last month
- 71