PV-Bench / experiments /t2 /task.toml
cielseven's picture
regenerate groundtruth with qcp 2.1.4, update the experiment harnesses
2504377
Raw History Blame Contribute Delete
1.21 kB
# Provisional limits; to be calibrated from gold builds (see notes/container-harness-design.md §14).
[task]
name = "t2"
agent_cwd = "/work"
prompt_file = "PROMPT.md"
tools_in_container = ["symexec.sh"]
eval_recipe = "t2"
eval_set_version = "v1"
runners = ["codex", "claude"]
egress_allow = ["api.openai.com", "chatgpt.com", "auth.openai.com", "api.anthropic.com"]
[limits]
memory = "8g"
cpus = 2
pids = 512
agent_timeout_s = 3600
eval_timeout_s = 3600 # P080: symexec alone is 7.5 min on the host; P081: 367 assumption probes.
check_timeout_s = 300 # judge session: cap on each compilation of the judge's own files in in-session and
# round checks (honest proofs seen so far: at most 155 s); the final check uses eval_timeout_s.
[mounts]
problem = { at = "/work/problem", mode = "ro" }
examples = { at = "/work/examples", mode = "ro" }
answer = { at = "/work/answer", mode = "rw" }
[judge]
# complete mount set for the judge container: answer/ is the judge's writable area
mounts = { workspace = { at = "/work/workspace", mode = "ro" }, answer = { at = "/work/answer", mode = "rw" }, harness = { at = "/work/harness", mode = "ro" }, gold = { at = "/work/gold", mode = "ro" } }