Add files using upload-large-folder tool

tasks/0001_042_1042725_qa_5/instruction.md

@@ -0,0 +1,17 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- marathon_results_2016.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which two countries have the most top 100 male marathon runners after the USA in the dataset?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer as a comma-separated list of the two country names.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_042_1042725_qa_5/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0001_042_1042725_qa_5"
+description = "Which two countries have the most top 100 male marathon runners after the USA in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/042/1042725.ipynb_qa_5"
+kaggle_dataset_name = "rojour/boston-results"
+gold_answer = "Kenya, Ethiopia"
+reward_mode_initial = "list"
+package_tier = 3
+difficulty_level = 2
+difficulty_tier = "medium"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "rojour__boston-results"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "rojour/boston-results"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "Kenya, Ethiopia"
+QUESTION = "Which two countries have the most top 100 male marathon runners after the USA in the dataset?"
+REWARD_MODE = "list"
+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]

tasks/0001_090_1090499_qa_1/instruction.md

@@ -0,0 +1,17 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- xAPI-Edu-Data.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which two nationalities have the highest representation in the dataset based on the analysis?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer as a comma-separated list of the two nationalities, in the order given.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_090_1090499_qa_1/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0001_090_1090499_qa_1"
+description = "Which two nationalities have the highest representation in the dataset based on the analysis?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/090/1090499.ipynb_qa_1"
+kaggle_dataset_name = "aljarah/xAPI-Edu-Data"
+gold_answer = "Kuwait, Jordan"
+reward_mode_initial = "list"
+package_tier = 1
+difficulty_level = 1
+difficulty_tier = "easy"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "aljarah__xAPI-Edu-Data"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "aljarah/xAPI-Edu-Data"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "Kuwait, Jordan"
+QUESTION = "Which two nationalities have the highest representation in the dataset based on the analysis?"
+REWARD_MODE = "list"
+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]

tasks/0001_155_1155264_qa_5/instruction.md

@@ -0,0 +1,15 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- (see /home/user/input)
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the most common instance type in the south zone identified through the analysis?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_155_1155264_qa_5/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-eval-v1/0001_155_1155264_qa_5"
+description = "What is the most common instance type in the south zone identified through the analysis?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/155/1155264.ipynb_qa_5"
+kaggle_dataset_name = "noqcks/aws-spot-pricing-market"
+gold_answer = "m4.large"
+reward_mode_initial = "exact_short"
+package_tier = 0
+difficulty_level = 2
+difficulty_tier = "medium"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "noqcks__aws-spot-pricing-market"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "noqcks/aws-spot-pricing-market"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "m4.large"
+QUESTION = "What is the most common instance type in the south zone identified through the analysis?"
+REWARD_MODE = "exact_short"
+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]

tasks/0001_257_1257756_qa_1/instruction.md

@@ -0,0 +1,17 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- FMEL_Dataset.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the average percentage of matches won by the home team across all seasons in the dataset?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Express the value as a percentage (e.g. 95.5), not a fraction.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_257_1257756_qa_1/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0001_257_1257756_qa_1"
+description = "What is the average percentage of matches won by the home team across all seasons in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/257/1257756.ipynb_qa_1"
+kaggle_dataset_name = "ricardomoya/football-matches-of-spanish-league"
+gold_answer = "51.16"
+reward_mode_initial = "numeric"
+package_tier = 1
+difficulty_level = 2
+difficulty_tier = "medium"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "ricardomoya__football-matches-of-spanish-league"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "ricardomoya/football-matches-of-spanish-league"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "51.16"
+QUESTION = "What is the average percentage of matches won by the home team across all seasons in the dataset?"
+REWARD_MODE = "flexible"
+ATOL = "0.05"
+RTOL = "0.01"
+
+[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]

tasks/0001_277_1277058_qa_5/instruction.md

@@ -0,0 +1,15 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- us_companies.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which year had the second-highest number of companies founded, and how many companies were founded that year?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_277_1277058_qa_5/task.toml

@@ -0,0 +1,58 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-eval-v1/0001_277_1277058_qa_5"
+description = "Which year had the second-highest number of companies founded, and how many companies were founded that year?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/277/1277058.ipynb_qa_5"
+kaggle_dataset_name = "govlab/open-data-500-companies"
+gold_answer = "2010 and 50"
+reward_mode_initial = "exact_short"
+package_tier = 1
+difficulty_level = 2
+difficulty_tier = "medium"
+
+[environment]
+build_timeout_sec = 600.0
+os = "linux"
+cpus = 2
+memory_mb = 4096
+storage_mb = 10240
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "govlab__open-data-500-companies"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "govlab/open-data-500-companies"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "2010 and 50"
+QUESTION = "Which year had the second-highest number of companies founded, and how many companies were founded that year?"
+REWARD_MODE = "exact_short"
+
+[agent]
+timeout_sec = 900.0
+
+[solution.env]

tasks/0001_312_1312239_qa_2/instruction.md

@@ -0,0 +1,15 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- 2015.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which region has the highest average Happiness Score when grouping by geographic regions?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_312_1312239_qa_2/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0001_312_1312239_qa_2"
+description = "Which region has the highest average Happiness Score when grouping by geographic regions?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/312/1312239.ipynb_qa_2"
+kaggle_dataset_name = "unsdsn/world-happiness"
+gold_answer = "Australia and New Zealand"
+reward_mode_initial = "exact_short"
+package_tier = 1
+difficulty_level = 2
+difficulty_tier = "medium"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "unsdsn__world-happiness"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "unsdsn/world-happiness"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "Australia and New Zealand"
+QUESTION = "Which region has the highest average Happiness Score when grouping by geographic regions?"
+REWARD_MODE = "exact_short"
+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]

tasks/0001_330_1330281_qa_3/instruction.md

@@ -0,0 +1,15 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- shot_logs.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the base shot success rate across all shots in the dataset, regardless of contextual factors?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_330_1330281_qa_3/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0001_330_1330281_qa_3"
+description = "What is the base shot success rate across all shots in the dataset, regardless of contextual factors?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/330/1330281.ipynb_qa_3"
+kaggle_dataset_name = "dansbecker/nba-shot-logs"
+gold_answer = "45.2139"
+reward_mode_initial = "numeric"
+package_tier = 1
+difficulty_level = 1
+difficulty_tier = "easy"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "dansbecker__nba-shot-logs"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "dansbecker/nba-shot-logs"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "45.2139"
+QUESTION = "What is the base shot success rate across all shots in the dataset, regardless of contextual factors?"
+REWARD_MODE = "numeric"
+ATOL = "0.001"
+RTOL = "0.005"
+
+[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]

tasks/0001_330_1330281_qa_4/instruction.md

@@ -0,0 +1,17 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- shot_logs.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+How does the 1st period’s shot success rate compare to the 4th period’s shot success rate during regulation?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer as: <period>, <percentage> pairs, comma-separated, with the period label first and the percentage as a plain number with two decimals followed by a percent sign (e.g., 46.05%).
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_330_1330281_qa_4/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0001_330_1330281_qa_4"
+description = "How does the 1st period’s shot success rate compare to the 4th period’s shot success rate during regulation?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/330/1330281.ipynb_qa_4"
+kaggle_dataset_name = "dansbecker/nba-shot-logs"
+gold_answer = "1st period 46.0528%, 4th period 44.0099%"
+reward_mode_initial = "list"
+package_tier = 1
+difficulty_level = 2
+difficulty_tier = "medium"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "dansbecker__nba-shot-logs"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "dansbecker/nba-shot-logs"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "1st period 46.0528%, 4th period 44.0099%"
+QUESTION = "How does the 1st period’s shot success rate compare to the 4th period’s shot success rate during regulation?"
+REWARD_MODE = "list"
+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]

tasks/0001_352_1352372_qa_5/instruction.md

@@ -0,0 +1,15 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- celebrity_deaths_4.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+How many celebrities in the dataset died as a result of accidents?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_352_1352372_qa_5/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0001_352_1352372_qa_5"
+description = "How many celebrities in the dataset died as a result of accidents?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/352/1352372.ipynb_qa_5"
+kaggle_dataset_name = "hugodarwood/celebrity-deaths"
+gold_answer = "141"
+reward_mode_initial = "numeric"
+package_tier = 1
+difficulty_level = 2
+difficulty_tier = "medium"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "hugodarwood__celebrity-deaths"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "hugodarwood/celebrity-deaths"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "141"
+QUESTION = "How many celebrities in the dataset died as a result of accidents?"
+REWARD_MODE = "numeric"
+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]

tasks/0001_361_1361614_qa_1/instruction.md

@@ -0,0 +1,15 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- GlobalLandTemperaturesByMajorCity.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which Indian city has the highest temperature difference between its maximum and minimum recorded temperatures in the dataset?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_361_1361614_qa_1/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0001_361_1361614_qa_1"
+description = "Which Indian city has the highest temperature difference between its maximum and minimum recorded temperatures in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/361/1361614.ipynb_qa_1"
+kaggle_dataset_name = "berkeleyearth/climate-change-earth-surface-temperature-data"
+gold_answer = "New Delhi"
+reward_mode_initial = "exact_short"
+package_tier = 1
+difficulty_level = 3
+difficulty_tier = "medium"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "berkeleyearth__climate-change-earth-surface-temperature-data"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "berkeleyearth/climate-change-earth-surface-temperature-data"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "New Delhi"
+QUESTION = "Which Indian city has the highest temperature difference between its maximum and minimum recorded temperatures in the dataset?"
+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]

tasks/0001_445_1445407_qa_4/instruction.md

@@ -0,0 +1,17 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- actual.csv
+- data_set_ALL_AML_independent.csv
+- data_set_ALL_AML_train.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the total number of patients in the test set used for evaluating the KNN model's performance?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_445_1445407_qa_4/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0001_445_1445407_qa_4"
+description = "What is the total number of patients in the test set used for evaluating the KNN model's performance?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/445/1445407.ipynb_qa_4"
+kaggle_dataset_name = "crawford/gene-expression"
+gold_answer = "34"
+reward_mode_initial = "numeric"
+package_tier = 1
+difficulty_level = 2
+difficulty_tier = "medium"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "crawford__gene-expression"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "crawford/gene-expression"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "34"
+QUESTION = "What is the total number of patients in the test set used for evaluating the KNN model's performance?"
+REWARD_MODE = "numeric"
+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]

tasks/0001_452_1452536_qa_1/instruction.md

@@ -0,0 +1,17 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- database.sqlite
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which team had the lowest True Performance in the dataset, and what was their True Performance value?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer as: <team>, <value> (comma-separated, team name first, keep the negative sign and decimals).
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_452_1452536_qa_1/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-eval-v1/0001_452_1452536_qa_1"
+description = "Which team had the lowest True Performance in the dataset, and what was their True Performance value?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/452/1452536.ipynb_qa_1"
+kaggle_dataset_name = "hugomathien/soccer"
+gold_answer = "Borussia Dortmund in the 2014/2015 season with -24.03 points."
+reward_mode_initial = "flexible"
+package_tier = 1
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "hugomathien__soccer"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "hugomathien/soccer"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "St. Mirren, with a True Performance value of -38.33"
+QUESTION = "Which team had the lowest True Performance in the dataset, and what was their True Performance value?"
+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]

tasks/0001_532_1532619_qa_4/instruction.md

@@ -0,0 +1,15 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- indian_liver_patient.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the difference in ROC AUC scores between the optimized SVM model (0.577) and the optimized Random Forest model (0.692) on the test dataset?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_532_1532619_qa_4/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0001_532_1532619_qa_4"
+description = "What is the difference in ROC AUC scores between the optimized SVM model (0.577) and the optimized Random Forest model (0.692) on the test dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/532/1532619.ipynb_qa_4"
+kaggle_dataset_name = "uciml/indian-liver-patient-records"
+gold_answer = "0.115"
+reward_mode_initial = "numeric"
+package_tier = 1
+difficulty_level = 1
+difficulty_tier = "easy"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "uciml__indian-liver-patient-records"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "uciml/indian-liver-patient-records"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "0.115"
+QUESTION = "What is the difference in ROC AUC scores between the optimized SVM model (0.577) and the optimized Random Forest model (0.692) on the test dataset?"
+REWARD_MODE = "numeric"
+ATOL = "0.001"
+RTOL = "0.005"
+
+[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]

tasks/0001_593_1593034_qa_5/instruction.md

@@ -0,0 +1,15 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- wineQualityReds.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the mean alcohol content of all wines in the dataset?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_593_1593034_qa_5/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0001_593_1593034_qa_5"
+description = "What is the mean alcohol content of all wines in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/593/1593034.ipynb_qa_5"
+kaggle_dataset_name = "piyushgoyal443/red-wine-dataset"
+gold_answer = "10.422983"
+reward_mode_initial = "numeric"
+package_tier = 1
+difficulty_level = 1
+difficulty_tier = "easy"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "piyushgoyal443__red-wine-dataset"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "piyushgoyal443/red-wine-dataset"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "10.422983"
+QUESTION = "What is the mean alcohol content of all wines in the dataset?"
+REWARD_MODE = "numeric"
+ATOL = "0.001"
+RTOL = "0.005"
+
+[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]

tasks/0001_636_1636611_qa_2/instruction.md

@@ -0,0 +1,16 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- iclr2017_papers.csv
+- iclr2017_conversations.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the most commonly assigned review rating in the ICLR 2017 dataset?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_636_1636611_qa_2/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0001_636_1636611_qa_2"
+description = "What is the most commonly assigned review rating in the ICLR 2017 dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/636/1636611.ipynb_qa_2"
+kaggle_dataset_name = "ahmaurya/iclr2017reviews"
+gold_answer = "6: Marginally above acceptance threshold"
+reward_mode_initial = "exact_short"
+package_tier = 0
+difficulty_level = 1
+difficulty_tier = "easy"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "ahmaurya__iclr2017reviews"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "ahmaurya/iclr2017reviews"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "6: Marginally above acceptance threshold"
+QUESTION = "What is the most commonly assigned review rating in the ICLR 2017 dataset?"
+REWARD_MODE = "exact_short"
+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]

tasks/0001_656_1656907_qa_2/instruction.md

@@ -0,0 +1,15 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- DigiDB_digimonlist.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the average (mean) Level 50 Attack value for all Digimon in the dataset?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_656_1656907_qa_2/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0001_656_1656907_qa_2"
+description = "What is the average (mean) Level 50 Attack value for all Digimon in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/656/1656907.ipynb_qa_2"
+kaggle_dataset_name = "rtatman/digidb"
+gold_answer = "124.52"
+reward_mode_initial = "numeric"
+package_tier = 0
+difficulty_level = 1
+difficulty_tier = "easy"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "rtatman__digidb"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "rtatman/digidb"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "124.52"
+QUESTION = "What is the average (mean) Level 50 Attack value for all Digimon in the dataset?"
+REWARD_MODE = "numeric"
+ATOL = "0.05"
+RTOL = "0.01"
+
+[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]

tasks/0001_656_1656907_qa_5/instruction.md

@@ -0,0 +1,15 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- DigiDB_digimonlist.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the average number of Equip Slots for all Digimon in the dataset?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_656_1656907_qa_5/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0001_656_1656907_qa_5"
+description = "What is the average number of Equip Slots for all Digimon in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/656/1656907.ipynb_qa_5"
+kaggle_dataset_name = "rtatman/digidb"
+gold_answer = "1.57"
+reward_mode_initial = "numeric"
+package_tier = 0
+difficulty_level = 1
+difficulty_tier = "easy"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "rtatman__digidb"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "rtatman/digidb"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "1.57"
+QUESTION = "What is the average number of Equip Slots for all Digimon in the dataset?"
+REWARD_MODE = "numeric"
+ATOL = "0.05"
+RTOL = "0.01"
+
+[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]

tasks/0001_667_1667317_qa_1/instruction.md

@@ -0,0 +1,15 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- camera_dataset.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the maximum zoom range (difference between maximum Zoom tele and minimum Zoom wide) in the dataset?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_667_1667317_qa_1/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0001_667_1667317_qa_1"
+description = "What is the maximum zoom range (difference between maximum Zoom tele and minimum Zoom wide) in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/667/1667317.ipynb_qa_1"
+kaggle_dataset_name = "crawford/1000-cameras-dataset"
+gold_answer = "518.0"
+reward_mode_initial = "numeric"
+package_tier = 1
+difficulty_level = 2
+difficulty_tier = "medium"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "crawford__1000-cameras-dataset"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "crawford/1000-cameras-dataset"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "518.0"
+QUESTION = "What is the maximum zoom range (difference between maximum Zoom tele and minimum Zoom wide) in the dataset?"
+REWARD_MODE = "numeric"
+ATOL = "0.05"
+RTOL = "0.01"
+
+[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]

tasks/0001_674_1674081_qa_1/instruction.md

@@ -0,0 +1,15 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- carInsurance_train.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the maximum value in the 'Balance' field before outlier removal in the dataset?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_674_1674081_qa_1/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0001_674_1674081_qa_1"
+description = "What is the maximum value in the 'Balance' field before outlier removal in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/674/1674081.ipynb_qa_1"
+kaggle_dataset_name = "kondla/carinsurance"
+gold_answer = "98417"
+reward_mode_initial = "numeric"
+package_tier = 1
+difficulty_level = 1
+difficulty_tier = "easy"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "kondla__carinsurance"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "kondla/carinsurance"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "98417"
+QUESTION = "What is the maximum value in the 'Balance' field before outlier removal in the dataset?"
+REWARD_MODE = "numeric"
+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]

tasks/0001_736_1736876_qa_4/instruction.md

@@ -0,0 +1,18 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- multipleChoiceResponses.csv
+- conversionRates.csv
+- schema.csv
+- freeformResponses.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+How many numeric columns are present in the dataset before additional processing?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_736_1736876_qa_4/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0001_736_1736876_qa_4"
+description = "How many numeric columns are present in the dataset before additional processing?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/736/1736876.ipynb_qa_4"
+kaggle_dataset_name = "kaggle/kaggle-survey-2017"
+gold_answer = "13"
+reward_mode_initial = "numeric"
+package_tier = 3
+difficulty_level = 1
+difficulty_tier = "easy"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "kaggle__kaggle-survey-2017"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "kaggle/kaggle-survey-2017"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "13"
+QUESTION = "How many numeric columns are present in the dataset before additional processing?"
+REWARD_MODE = "numeric"
+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]

tasks/0001_737_1737901_qa_1/instruction.md

@@ -0,0 +1,16 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- multipleChoiceResponses.csv
+- freeformResponses.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the median age of respondents in the dataset?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_737_1737901_qa_1/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0001_737_1737901_qa_1"
+description = "What is the median age of respondents in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/737/1737901.ipynb_qa_1"
+kaggle_dataset_name = "kaggle/kaggle-survey-2017"
+gold_answer = "30"
+reward_mode_initial = "numeric"
+package_tier = 1
+difficulty_level = 1
+difficulty_tier = "easy"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "kaggle__kaggle-survey-2017"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "kaggle/kaggle-survey-2017"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "30"
+QUESTION = "What is the median age of respondents in the dataset?"
+REWARD_MODE = "numeric"
+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]

tasks/0001_741_1741634_qa_2/instruction.md

@@ -0,0 +1,16 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- multipleChoiceResponses.csv
+- conversionRates.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+After excluding compensation outliers (values below $1 or above $2,000,000), what is the maximum compensation value retained in the dataset for analysis?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_741_1741634_qa_2/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0001_741_1741634_qa_2"
+description = "After excluding compensation outliers (values below $1 or above $2,000,000), what is the maximum compensation value retained in the dataset for analysis?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/741/1741634.ipynb_qa_2"
+kaggle_dataset_name = "kaggle/kaggle-survey-2017"
+gold_answer = "2000000"
+reward_mode_initial = "numeric"
+package_tier = 0
+difficulty_level = 2
+difficulty_tier = "medium"
+
+[environment]
+build_timeout_sec = 600.0
+os = "linux"
+cpus = 1
+memory_mb = 1024
+storage_mb = 5120
+gpus = 0
+allow_internet = true
+mcp_servers = []
+
+# Pre-agent hook: Harbor runs the command AFTER container start and BEFORE the
+# agent setup begins. We use it to pull this task's bucket prefix into
+# /home/user/input/. See environment/pull_bucket.py.
+[environment.healthcheck]
+command = "python3 /opt/pull_bucket.py && [ -n \"$(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 = "kaggle__kaggle-survey-2017"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "kaggle/kaggle-survey-2017"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "2000000"
+QUESTION = "After excluding compensation outliers (values below $1 or above $2,000,000), what is the maximum compensation value retained in the dataset for analysis?"
+REWARD_MODE = "numeric"
+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]

tasks/0001_761_1761668_qa_2/instruction.md

@@ -0,0 +1,17 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- data_set_ALL_AML_independent.csv
+- data_set_ALL_AML_train.csv
+- actual.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+How many columns containing the term "call" were removed from the training dataset during preprocessing?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_761_1761668_qa_2/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0001_761_1761668_qa_2"
+description = "How many columns containing the term \"call\" were removed from the training dataset during preprocessing?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/761/1761668.ipynb_qa_2"
+kaggle_dataset_name = "crawford/gene-expression"
+gold_answer = "38"
+reward_mode_initial = "numeric"
+package_tier = 1
+difficulty_level = 2
+difficulty_tier = "medium"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "crawford__gene-expression"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "crawford/gene-expression"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "38"
+QUESTION = "How many columns containing the term \"call\" were removed from the training dataset during preprocessing?"
+REWARD_MODE = "numeric"
+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]

tasks/0001_761_1761668_qa_3/instruction.md

@@ -0,0 +1,15 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- (see /home/user/input)
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the median value of the gene expression for the gene 'X64594_at' in the training dataset's sample subset?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_761_1761668_qa_3/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-eval-v1/0001_761_1761668_qa_3"
+description = "What is the median value of the gene expression for the gene 'X64594_at' in the training dataset's sample subset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/761/1761668.ipynb_qa_3"
+kaggle_dataset_name = "crawford/gene-expression"
+gold_answer = "-134.0"
+reward_mode_initial = "numeric"
+package_tier = 0
+difficulty_level = 2
+difficulty_tier = "medium"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "crawford__gene-expression"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "crawford/gene-expression"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "-134.0"
+QUESTION = "What is the median value of the gene expression for the gene 'X64594_at' in the training dataset's sample subset?"
+REWARD_MODE = "numeric"
+ATOL = "0.05"
+RTOL = "0.01"
+
+[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]

tasks/0001_869_1869604_qa_2/instruction.md

@@ -0,0 +1,15 @@
+You are a data-analysis agent working in a sandbox. Use your code-execution tool to inspect the files and compute the answer.
+
+Files (in /home/user/input, no subfolders):
+- creditcard.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the average transaction amount across all records in the dataset?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer with a single clean value: a bare number (no commas or units, e.g. 95293), a short label, yes/no, or a comma-separated list. Keep decimal precision. If there's no applicable answer, write: Not Applicable
+
+Write only that value to /workdir/answer.txt (e.g. `echo -n "<value>" > /workdir/answer.txt`), then stop.

tasks/0001_869_1869604_qa_2/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0001_869_1869604_qa_2"
+description = "What is the average transaction amount across all records in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/869/1869604.ipynb_qa_2"
+kaggle_dataset_name = "mlg-ulb/creditcardfraud"
+gold_answer = "88.35"
+reward_mode_initial = "numeric"
+package_tier = 0
+difficulty_level = 1
+difficulty_tier = "easy"
+
+[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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "mlg-ulb__creditcardfraud"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "mlg-ulb/creditcardfraud"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "88.35"
+QUESTION = "What is the average transaction amount across all records in the dataset?"
+REWARD_MODE = "numeric"
+ATOL = "0.05"
+RTOL = "0.01"
+
+[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]