schema_version = "1.2"
artifacts = []

[task]
name = "train-verify/0001_181_1181828_qa_4"
description = "Which model demonstrated the highest training accuracy but the lowest test accuracy in the comparison analysis?"
authors = []
keywords = ["data-agent", "data-analysis", "kaggle"]

[metadata]
source_dataset = "jupyter-agent/jupyter-agent-dataset"
source_row_id = "0001/181/1181828.ipynb_qa_4"
kaggle_dataset_name = "uciml/glass"
gold_answer = "Decision Tree"
reward_mode_initial = "exact_short"
package_tier = 2
difficulty_level = 4
difficulty_tier = "hard"

[environment]
build_timeout_sec = 600.0
os = "linux"
cpus = 1
memory_mb = 1024
storage_mb = 5120
gpus = 0
allow_internet = true
mcp_servers = []

# Pre-agent hook: Harbor runs the command AFTER container start and BEFORE the
# agent setup begins. We use it to pull this task's bucket prefix into
# /home/user/input/. See environment/pull_bucket.py.
[environment.healthcheck]
command = "python3 /opt/pull_bucket.py && [ -n \"$(ls /home/user/input)\" ]"
interval_sec = 2.0
timeout_sec = 180.0
start_period_sec = 5.0
start_interval_sec = 2.0
retries = 30

[environment.env]
HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
BUCKET_PREFIX = "uciml__glass"
HF_TOKEN = "${HF_TOKEN}"
KAGGLE_DATASET_NAME = "uciml/glass"

[verifier]
timeout_sec = 120.0

[verifier.env]
EXPECTED_ANSWER = "Decision Tree"
QUESTION = "Which model demonstrated the highest training accuracy but the lowest test accuracy in the comparison analysis?"
REWARD_MODE = "flexible"
ATOL = "0.0"
RTOL = "0.0"

[agent]
# Capped at 600s (10 min) to kill the long-tail stuck-agent cases without
# cutting off legitimate complex trials. Median Phase B trial is 60-120s;
# legitimate L4/L5 tasks can hit 200-300s; anything past 600s is almost
# certainly a stuck agent loop.
timeout_sec = 600.0

[solution.env]