Open benchmark harness for comparing current major AI models on general work: reasoning, coding, structured output, long-context extraction, tool-use readiness, and lightweight data analysis.

This repository is a reproducibility-first benchmark starter. It publishes the model roster, task format, scoring code, and raw result schema before any leaderboard claims.

Roster snapshot: 2026年05月05日.

The benchmark roster is in model_roster.json. It currently tracks OpenAI, Anthropic, Google, xAI, DeepSeek, Mistral, Cohere, Alibaba Qwen, Moonshot Kimi, MiniMax, Z.AI GLM, and Meta Llama.

Provider model names move quickly. Every run should record:

• exact model ID used by the provider
• whether the ID is a stable snapshot, alias, preview, or partner-hosted ID
• temperature, max output, reasoning mode, tool access, and timestamp
• raw prompt and raw output

See docs/research-notes.md for the source notes behind the initial roster.

The public seed suite in tasks/general_seed.jsonl is intentionally small. It validates the harness and scoring path. Larger benchmark suites should be versioned under tasks/vYYYY-MM-DD/ and should include public seed tasks plus private holdout tasks.

Task families:

• reasoning: compact logic and quantitative reasoning
• coding: bug diagnosis and patch generation
• structured-output: JSON extraction and schema discipline
• long-context: retrieval of the exact relevant clause from distractors
• data-analysis: table arithmetic and probabilistic scoring
• calibration: probability math and uncertainty discipline

Validate roster/tasks and emit placeholder rows:

python src/sf_benchmark/runner.py \
 --models model_roster.json \
 --tasks tasks/general_seed.jsonl \
 --out results/dry-run.json

Score a JSONL file of model responses:

python src/sf_benchmark/runner.py \
 --models model_roster.json \
 --tasks tasks/general_seed.jsonl \
 --responses examples/responses.seed.jsonl \
 --out results/scored-seed.json

Run tests:

python -m unittest discover -s tests

--responses expects JSONL:

{"model":"openai/gpt-5.5","task_id":"data.table.001","output":"0.715"}

Results follow schemas/result.schema.json.

• Use the same prompt and same tool policy across models in the same run.
• Prefer stable snapshot IDs; if using aliases, record the alias resolution date.
• Do not rank models until raw outputs, scoring code, and run metadata are public.
• Report accuracy next to latency, cost, refusal/error rate, and output length.
• Keep public seed tasks separate from private holdout tasks to reduce benchmark overfitting.

The methodology is in docs/methodology.md.

The companion domain benchmark is spfunctions/prediction-market-model-benchmark.

MIT.