AeroAIAviation/AvBench1.0_BENCHMARK

Offline Aviation Model Benchmark

This folder contains an offline benchmark for comparing local Ollama models and OpenAI-compatible local endpoints such as Unsloth Studio. It covers 10 aviation topics with 40 cases: 20 calculation/procedure prompts and 20 short-answer knowledge questions.

All route, weather, aircraft, and chart data are fictional or explicitly supplied in the prompts. This benchmark is for model comparison and education; it is not operational flight-planning guidance and is not checked against current FAA publications. No web search is used by the benchmark.

How the benchmark works

1. benchmark.py loads cases from benchmark.json and sends each prompt to the selected local model.
2. If a case has an image field, the runner also sends that image. The visual chart case uses assets/visual_chart.png.
3. The grader checks the answer for expected numbers within the case's declared tolerance and for required case-insensitive text terms.
4. It writes the prompt, answer, per-check outcomes, and timing to JSON and CSV, plus a Markdown report with overall, category, and question-style scores.

The automated grade is a checklist score, not an expert assessment. rubric entries in benchmark.json are human-review guidance and do not add points to the automated score. The term checks are literal substrings, so valid synonyms may be marked as misses. Review raw answers before drawing conclusions.

Ollama requests are capped at 768 generated tokens and ask for think: false where supported, so a reasoning trace does not consume the answer budget. OpenAI-compatible calls are capped at 768 tokens. Python tool calls are off by default. If enabled, model-generated Python code runs locally without a security sandbox; enable it only for models/prompts you trust.

Run it

Requirements: Python 3.9+ and a running local model server. The scripts use Python's standard library and require no package installation.

The current models.json is configured for the Gemma tag used in the latest run. To run all 40 cases directly:

Or run the model configured in models.json through the matrix launcher:

python3 run_matrix.py --config models.json

Run only the short-answer cases:

python3 benchmark.py --provider ollama --model gemma4:latest --mode qa

Useful filters:

python3 benchmark.py --provider ollama --model gemma4:latest --category "Aerospace Engineering"
python3 benchmark.py --provider ollama --model gemma4:latest --mode qa --category "Aerospace Engineering"
python3 benchmark.py --provider ollama --model gemma4:latest --limit 8 --seed 42

For Unsloth Studio, start its OpenAI-compatible local server, then use its actual model name and endpoint URL:

python3 benchmark.py --provider openai --base-url http://localhost:3000/v1 --model YOUR_MODEL_NAME

To compare endpoints, edit models.json to include one object per model, then run python3 run_matrix.py --config models.json. Use models.example.json as a starting point. Add --mode qa or --category "..." to the matrix command to filter the cases.

Optional Python-tool run:

python3 benchmark.py --provider ollama --model gemma4:latest --enable-python

The --enable-python flag executes code generated by the model in a local subprocess. It is not sandboxed. Leave the flag off for a normal run.

Files and resources

File or folder	Purpose
README.md	Human run guide and scoring explanation
CLAUDE.md	Project context and safe editing guidance for Claude Code
benchmark.py	Single-model runner, local API clients, grader, and report writer
run_matrix.py	Runs each configured model sequentially and creates a comparison report
benchmark.json	All 40 prompts, expected checks, category labels, rubrics, and image references
assets/visual_chart.png	Synthetic chart supplied to multimodal visual-reading cases
models.json	Active local endpoint/model configuration; ignored by Git
models.example.json	Example Ollama and Unsloth configuration
results/	Timestamped model answers, scores, and reports; ignored by Git
.gitignore	Excludes local config, results, and Python cache files from Git

Each run gets its own directory inside results/, containing results.json, results.csv, and report.md. The JSON is the detailed record; the Markdown report is the quick summary. Runs do not overwrite each other. If you stop a run, completed cases remain saved, but the runner does not resume automatically.

Existing runs

• results/20261007T210421Z-gemma4_latest/ is the Gemma 4 run stopped at the user's request after 28 of 40 cases. Its report explicitly marks the scores as partial.
• Other timestamped directories contain partial Qwen runs from earlier attempts. Check the model, status, and completed case count in results.json before comparing runs.

Editing the dataset

Add or update cases in benchmark.json. Each case needs a unique id, category, mode, prompt, and checks. Optional fields include image, rubric, and tool_preferred.

• mode is text, image, or qa; it enables filtering with --mode.
• checks.numbers contains { "value": ..., "tolerance": ... } checks. A check passes if any number in the model answer is within tolerance of the target.
• checks.terms contains literal case-insensitive substrings that must appear in the answer.
• rubric is for human review only.
• tool_preferred documents which cases may benefit from --enable-python; it does not activate the tool by itself.

Preserve offline operation and keep examples fictional/supplied-data-based. Do not treat benchmark outputs as validated real-world aviation advice.