VAHBench / benchmark /config /default.yaml
hobezhang's picture
Curate five benchmark prompts and 15 GitHub-aligned showcase runs
3595fbe verified
Raw History Blame Contribute Delete
2.55 kB
# Everything the evaluator needs that is not "which data, where to write, how
# thorough". Those three stay on the command line; the rest lives here so that a
# run is described by a file you can archive next to its report, rather than by a
# command line nobody wrote down.
#
# Override any single value for one run with an environment variable, e.g.
# VAB_EVAL_JOBS=4 python run.py --data data
# Original measurement-module selection. Coverage records unavailable evidence.
# spec: delivered media specifications and source timing
# artifact: frozen/blank/flat frames and join artifacts
# visual: identity consistency and repeated footage
# physical: physical contact/penetration evidence
# av: audio clipping and dynamic range
# seam: cut continuity
modules: [spec, artifact, visual, physical, av, seam]
# When one of the paper's original detectors is not installed under detectors/
# (or VAB_EVAL_DETECTOR_ROOT), use the bundled reference reimplementation for
# that module. It is a different instrument: it never makes a run "full", it
# needs --allow-partial, and its numbers are not comparable with the paper's.
# There is none for `physical`; reference `visual` measures repeats only (no
# face identity) and reference `av` measures clipping only (no dynamic range).
reference_detectors: true
# Torch device for modules that use a model. CPU is the reference configuration.
device: cpu
# Scoring thresholds. Kept in its own file because its hash is archived with
# every report -- changing it changes the benchmark version.
cut_points: config/thresholds.yaml
# Parallel runs. Each run spawns per-module subprocesses with their own venvs,
# so this multiplies memory, not just CPU.
jobs: 1
# Trajectory fact extraction. The deterministic extractors always run; the model
# only fills what they could not reach.
lm:
provider: openrouter
# The extraction model is part of the benchmark configuration. Record any
# change with results; a provider model ID is not an immutable weight hash.
model: z-ai/glm-5.3
max_chars: 120000
# Output budget, shared by reasoning and answer tokens. Too small and the
# model returns an empty string that looks like an empty finding.
max_tokens: 32768
# null uses the provider's native variable (e.g. OPENROUTER_API_KEY).
# Set a custom variable name here, never the secret itself.
api_key_env: null
# Optional path to a file holding the key instead of an environment variable.
api_key_file: null
# Speech recognition model for the av module's transcript checks.
asr_model: base