schema_version = "1.1"

[task]
name = "apodex-ai/task-006-matbench-expt-gap-cleaning"
description = "清洗Matbench实验带隙数据并训练组成Ridge基线模型"
authors = [{ name = "Apodex AI & GADE Union" }]
keywords = ["materials-informatics", "matbench", "band-gap", "data-cleaning", "ridge-regression", "chemical-formula-parsing"]

[metadata]
frontierchallenge_task_id = "task_006_matbench_expt_gap_cleaning"
difficulty = "hard"
category = "真实开源数据处理；材料信息学；化学式解析；单位换算；重复记录聚合；机器学习基线；自动评分"

[agent]
user = "mambauser"
timeout_sec = 3600.0
network_mode = "public"

[verifier]
timeout_sec = 600.0

[verifier.env]
JUDGE_API_KEY = "${JUDGE_API_KEY}"
JUDGE_API_ENDPOINT = "${JUDGE_API_ENDPOINT}"
JUDGE_MODEL = "gpt-5.6-sol"
JUDGE_REASONING_EFFORT = "high"
JUDGE_REPEATS = "3"

[environment]
network_mode = "public"
build_timeout_sec = 1200.0
cpus = 2
memory_mb = 4096
storage_mb = 2048
gpus = 0

# LOCAL_SMOKE_ONLY: agent network changed to public because macOS Docker cannot enforce Harbor allowlists.