Configuration

Set openbench model and evaluation behavior

Configuration Methods

openbench can be configured through multiple methods, in order of precedence:

1. Command-Line Arguments

# Run eval with --config CLI flags
bench mmlu --max_tokens=10000 --temperature=0.5

See what else you can do with the openbench CLI.

2. Environment Variables

# Set configuration defaults
export BENCH_MAX_TOKENS=10000
export BENCH_TEMPERATURE=0.5

3. Configuration Files (.env files)

.env

# Write default values to config file
BENCH_MAX_TOKENS=10000
BENCH_TEMPERATURE=0.5

# Automatically loaded if in current directory
bench eval mmlu

# Or source manually
source .env
bench eval mmlu

If not defined through any of the above methods, parameters resort to default values.

Configuration Parameters

Model Configuration

CLI Command	Environment Variable	Description
--model	BENCH_MODEL	Model(s) to evaluate.
--model-base-url	BENCH_MODEL_BASE_URL	Base URL for model(s).
--model-role	BENCH_MODEL_ROLE	Map role(s) to specific models.

Some providers may support additional parameters. See provider-specific model configuration for extended features.

Task Configuration

CLI Command	Environment Variable	Description
--epochs	BENCH_EPOCHS	Number of epochs to run each evaluation.
--epochs-reducer	BENCH_EPOCHS_REDUCER	Reducer(s) applied when aggregating epoch scores.
--limit	BENCH_LIMIT	Limit evaluated samples (single number or range).
--message-limit	BENCH_MESSAGE_LIMIT	Maximum number of messages per sample.
--score / --no-score	BENCH_SCORE	Grade the benchmark, or skip scoring (can score later).

Use -T Flag for Other Task-Specific Arguments
(e.g. bench eval graphwalks --model ollama/llama3.1:70b -T task=parents)

Generation Settings

CLI Command	Environment Variable	Description
--temperature	BENCH_TEMPERATURE	Model sampling temperature.
--top-p	BENCH_TOP_P	Nucleus sampling.
--max-tokens	BENCH_MAX_TOKENS	Maximum tokens for model response.
--seed	BENCH_SEED	Random seed for deterministic generation.

Performance & Concurrency

CLI Command	Environment Variable	Description
--max-connections	BENCH_MAX_CONNECTIONS	Maximum number of parallel requests.
--max-subprocesses	BENCH_MAX_SUBPROCESSES	Maximum number of parallel subprocesses.
--max-tasks	BENCH_MAX_TASKS	Maximum number of tasks to run concurrently.
--timeout	BENCH_TIMEOUT	Timeout per model API request (seconds).
--max-retries	BENCH_MAX_RETRIES	Maximum number of retries for API requests.
--retry-on-error	BENCH_RETRY_ON_ERROR	Retry samples on errors (set number of retries).
--fail-on-error	BENCH_FAIL_ON_ERROR	Failure threshold for sample errors (percentage or count).
--no-fail-on-error	BENCH_NO_FAIL_ON_ERROR	Do not fail evaluation if errors occur.

Logging & Output

CLI Flag	Environment Variable	Description
--logfile	BENCH_OUTPUT	Output file for results.
--log-format	BENCH_LOG_FORMAT	Output logging format (eval / json).
--display	BENCH_DISPLAY	Display type for evaluation progress.
--log-samples / --no-log-samples	BENCH_LOG_SAMPLES	Include or exclude detailed samples in logs.
--log-images / --no-log-images	BENCH_LOG_IMAGES	Include or exclude base64 encoded images in logs.
--log-buffer	BENCH_LOG_BUFFER	Number of samples to buffer before writing logs.
--log-dir	BENCH_LOG_DIR	Directory for log files.

Sandbox & Execution

CLI Flag	Environment Variable	Description
--sandbox	BENCH_SANDBOX	Environment to run evaluation (local or docker).
--sandbox-cleanup / --no-sandbox-cleanup	BENCH_SANDBOX_CLEANUP	Cleanup sandbox environments after tasks (or skip cleanup).

Hub Integration

CLI Flag	Environment Variable	Description
--hub-repo	BENCH_HUB_REPO	Target Hub dataset repo for logs.
--hub-private	BENCH_HUB_PRIVATE	Push Hub dataset as private.

Debugging & Inspection

CLI Flag	Environment Variable	Description
--debug	BENCH_DEBUG	Enable debug mode with full stack traces.
--debug-errors	BENCH_DEBUG_ERRORS	Enable debug mode for errors only.
--trace	BENCH_TRACE	Trace message interactions with model.
--alpha	BENCH_ALPHA	Allow running experimental/alpha benchmarks.

Configuration Examples

Simple Custom Configuration

bench eval humaneval \
  --model openai/gpt-4o \
  --temperature 0.5 \
  --max-tokens 512 \
  --epochs 5 \
  --sandbox docker

Model Comparison

bench eval mmlu --limit 500 \
    --model groq/llama-3.3-70b \
    --model openai/gpt-4o \
    --model anthropic/claude-3-5-sonnet

Multi-Model Evaluation

bench eval simpleqa \
  --model-role candidate=groq/llama-3.3-70b \
  --model-role grader=openai/gpt-4o

Batch Configuration

bench eval --model openai/gpt-4.1 \
  math --limit 20 \
  mmlu --limit 20

Interruption Retry Configuration

bench eval-retry logs/incomplete.json \
  --max-retries 3 \
  --retry-on-error

Display Configuration

# Display modes
export BENCH_DISPLAY="rich"           # Rich terminal output (default)
export BENCH_DISPLAY="plain"          # Simple text output
export BENCH_DISPLAY="none"           # No output (logs only)
export BENCH_DISPLAY="conversation"   # Show full conversations

bench eval mmlu --display conversation