Add files using upload-large-folder tool

tasks/0001_604_1604140_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):
+- (see /home/user/input)
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which U.S. state has the highest number of recorded mass shootings between 2010 and 2017 according to 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_664_1664478_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):
+- cereal.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which manufacturer has the highest number of cereal products 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_664_1664478_qa_1/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0001_664_1664478_qa_1"
+description = "Which manufacturer has the highest number of cereal products in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/664/1664478.ipynb_qa_1"
+kaggle_dataset_name = "crawford/80-cereals"
+gold_answer = "K"
+reward_mode_initial = "exact_short"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "crawford__80-cereals"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "crawford/80-cereals"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "K"
+QUESTION = "Which manufacturer has the highest number of cereal products in the 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_840_1840182_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):
+- winemag-data-130k-v2.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the wine variety with the largest number of reviews 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_840_1840182_qa_4/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0001_840_1840182_qa_4"
+description = "What is the wine variety with the largest number of reviews in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/840/1840182.ipynb_qa_4"
+kaggle_dataset_name = "zynicide/wine-reviews"
+gold_answer = "Pinot Noir"
+reward_mode_initial = "exact_short"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "zynicide__wine-reviews"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "zynicide/wine-reviews"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "Pinot Noir"
+QUESTION = "What is the wine variety with the largest number of reviews in the 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_878_1878746_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):
+- cereal.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the maximum sodium content found in any cereal 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_878_1878746_qa_4/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0001_878_1878746_qa_4"
+description = "What is the maximum sodium content found in any cereal in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/878/1878746.ipynb_qa_4"
+kaggle_dataset_name = "crawford/80-cereals"
+gold_answer = "320"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "crawford__80-cereals"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "crawford/80-cereals"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "320"
+QUESTION = "What is the maximum sodium content found in any cereal 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_990_1990794_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):
+- 911_calls_for_service.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which priority level has the highest percentage of calls 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_990_1990794_qa_2/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0001_990_1990794_qa_2"
+description = "Which priority level has the highest percentage of calls in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0001/990/1990794.ipynb_qa_2"
+kaggle_dataset_name = "sohier/baltimore-911-calls"
+gold_answer = "Medium"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "sohier__baltimore-911-calls"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "sohier/baltimore-911-calls"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "Medium"
+QUESTION = "Which priority level has the highest percentage of calls in the 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/0010_191_10191909_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):
+- data.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the percentage of benign tumors 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/0010_191_10191909_qa_1/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0010_191_10191909_qa_1"
+description = "What is the percentage of benign tumors in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0010/191/10191909.ipynb_qa_1"
+kaggle_dataset_name = "uciml/breast-cancer-wisconsin-data"
+gold_answer = "62.7417"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "uciml__breast-cancer-wisconsin-data"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "uciml/breast-cancer-wisconsin-data"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "62.7417"
+QUESTION = "What is the percentage of benign tumors 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/0012_593_12593899_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):
+- crime.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which day of the month has the highest crime rate based on 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/0012_593_12593899_qa_2/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0012_593_12593899_qa_2"
+description = "Which day of the month has the highest crime rate based on the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0012/593/12593899.ipynb_qa_2"
+kaggle_dataset_name = "wosaku/crime-in-vancouver"
+gold_answer = "1"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "wosaku__crime-in-vancouver"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "wosaku/crime-in-vancouver"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "1"
+QUESTION = "Which day of the month has the highest crime rate based on 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/0014_390_14390255_qa_5/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):
+- student-mat.csv
+- student-por.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What percentage of students in the Portuguese dataset have parents who live together (Pstatus=T)?
+
+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/0014_390_14390255_qa_5/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0014_390_14390255_qa_5"
+description = "What percentage of students in the Portuguese dataset have parents who live together (Pstatus=T)?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0014/390/14390255.ipynb_qa_5"
+kaggle_dataset_name = "uciml/student-alcohol-consumption"
+gold_answer = "87.67"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "uciml__student-alcohol-consumption"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "uciml/student-alcohol-consumption"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "87.67"
+QUESTION = "What percentage of students in the Portuguese dataset have parents who live together (Pstatus=T)?"
+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/0016_945_16945752_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):
+- haberman.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the standard deviation of axillary node counts for patients who survived versus those who did not?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer as: <survived_std>, <not_survived_std> (comma-separated, plain numbers).
+
+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/0016_945_16945752_qa_5/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0016_945_16945752_qa_5"
+description = "What is the standard deviation of axillary node counts for patients who survived versus those who did not?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0016/945/16945752.ipynb_qa_5"
+kaggle_dataset_name = "gilsousa/habermans-survival-data-set"
+gold_answer = "5.86, 9.13"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "gilsousa__habermans-survival-data-set"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "gilsousa/habermans-survival-data-set"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "5.86, 9.13"
+QUESTION = "What is the standard deviation of axillary node counts for patients who survived versus those who did not?"
+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/0016_971_16971389_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):
+- mushrooms.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the most predictive feature for distinguishing edible from poisonous mushrooms based on value counts 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/0016_971_16971389_qa_1/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0016_971_16971389_qa_1"
+description = "What is the most predictive feature for distinguishing edible from poisonous mushrooms based on value counts analysis?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0016/971/16971389.ipynb_qa_1"
+kaggle_dataset_name = "uciml/mushroom-classification"
+gold_answer = "odor"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "uciml__mushroom-classification"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "uciml/mushroom-classification"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "odor"
+QUESTION = "What is the most predictive feature for distinguishing edible from poisonous mushrooms based on value counts 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/0017_291_17291057_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):
+- winequality-red.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which feature in the dataset shows the strongest positive correlation with wine quality according to the correlation matrix 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/0017_291_17291057_qa_1/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0017_291_17291057_qa_1"
+description = "Which feature in the dataset shows the strongest positive correlation with wine quality according to the correlation matrix analysis?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0017/291/17291057.ipynb_qa_1"
+kaggle_dataset_name = "uciml/red-wine-quality-cortez-et-al-2009"
+gold_answer = "alcohol"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "uciml__red-wine-quality-cortez-et-al-2009"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "uciml/red-wine-quality-cortez-et-al-2009"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "alcohol"
+QUESTION = "Which feature in the dataset shows the strongest positive correlation with wine quality according to the correlation matrix 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/0025_564_25564899_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):
+- gender-classifier-DFE-791531.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which model achieved the highest accuracy when using only the tweet text as input?
+
+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/0025_564_25564899_qa_4/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0025_564_25564899_qa_4"
+description = "Which model achieved the highest accuracy when using only the tweet text as input?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0025/564/25564899.ipynb_qa_4"
+kaggle_dataset_name = "crowdflower/twitter-user-gender-classification"
+gold_answer = "Logistic Regression"
+reward_mode_initial = "exact_short"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "crowdflower__twitter-user-gender-classification"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "crowdflower/twitter-user-gender-classification"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "Logistic Regression"
+QUESTION = "Which model achieved the highest accuracy when using only the tweet text as input?"
+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/0027_004_27004450_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):
+- games.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which opening for white has the highest win percentage among the top 10 most common openings 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/0027_004_27004450_qa_2/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-eval-v1/0027_004_27004450_qa_2"
+description = "Which opening for white has the highest win percentage among the top 10 most common openings in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0027/004/27004450.ipynb_qa_2"
+kaggle_dataset_name = "datasnaek/chess"
+gold_answer = "Philiodor Defense #3"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "datasnaek__chess"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "datasnaek/chess"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "Scandinavian Defense: Mieses-Kotroc Variation"
+QUESTION = "Which opening for white has the highest win percentage among the top 10 most common openings in the 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/0027_179_27179583_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):
+- HN_posts_year_to_Sep_26_2016.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 comments for Ask HN posts posted at 15:00?
+
+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/0027_179_27179583_qa_2/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0027_179_27179583_qa_2"
+description = "What is the average number of comments for Ask HN posts posted at 15:00?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0027/179/27179583.ipynb_qa_2"
+kaggle_dataset_name = "hacker-news/hacker-news-posts"
+gold_answer = "28.68"
+reward_mode_initial = "numeric"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "hacker-news__hacker-news-posts"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "hacker-news/hacker-news-posts"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "28.68"
+QUESTION = "What is the average number of comments for Ask HN posts posted at 15:00?"
+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/0027_503_27503967_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):
+- data.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What was the highest test set accuracy achieved by any of the models evaluated (kNN, Random Forest, Naive Bayes) using their optimized parameters?
+
+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/0027_503_27503967_qa_5/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0027_503_27503967_qa_5"
+description = "What was the highest test set accuracy achieved by any of the models evaluated (kNN, Random Forest, Naive Bayes) using their optimized parameters?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0027/503/27503967.ipynb_qa_5"
+kaggle_dataset_name = "uciml/breast-cancer-wisconsin-data"
+gold_answer = "0.95"
+reward_mode_initial = "numeric"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "uciml__breast-cancer-wisconsin-data"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "uciml/breast-cancer-wisconsin-data"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "0.95"
+QUESTION = "What was the highest test set accuracy achieved by any of the models evaluated (kNN, Random Forest, Naive Bayes) using their optimized parameters?"
+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/0029_184_29184728_qa_1/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0029_184_29184728_qa_1"
+description = "Which species of iris has the highest average sepal width according to the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0029/184/29184728.ipynb_qa_1"
+kaggle_dataset_name = "uciml/iris"
+gold_answer = "Iris-setosa"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "uciml__iris"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "uciml/iris"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "Iris-setosa"
+QUESTION = "Which species of iris has the highest average sepal width according to the 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/0029_630_29630344_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):
+- Online Retail.xlsx
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+How many transactions were removed as duplicates from the original 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/0029_630_29630344_qa_4/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0029_630_29630344_qa_4"
+description = "How many transactions were removed as duplicates from the original dataset during preprocessing?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0029/630/29630344.ipynb_qa_4"
+kaggle_dataset_name = "jihyeseo/online-retail-data-set-from-uci-ml-repo"
+gold_answer = "5268"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "jihyeseo__online-retail-data-set-from-uci-ml-repo"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "jihyeseo/online-retail-data-set-from-uci-ml-repo"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "5268"
+QUESTION = "How many transactions were removed as duplicates from the original 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/0031_074_31074287_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):
+- winemag-data-130k-v2.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the average vintage year of the wines in the dataset after removing entries with missing years?
+
+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/0031_074_31074287_qa_3/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0031_074_31074287_qa_3"
+description = "What is the average vintage year of the wines in the dataset after removing entries with missing years?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0031/074/31074287.ipynb_qa_3"
+kaggle_dataset_name = "zynicide/wine-reviews"
+gold_answer = "2010.67"
+reward_mode_initial = "numeric"
+package_tier = 2
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "zynicide__wine-reviews"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "zynicide/wine-reviews"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "2010.67"
+QUESTION = "What is the average vintage year of the wines in the dataset after removing entries with missing years?"
+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/0031_954_31954480_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):
+- Iris.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which species exhibits the highest average petal width according to the feature averages 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/0031_954_31954480_qa_3/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0031_954_31954480_qa_3"
+description = "Which species exhibits the highest average petal width according to the feature averages analysis?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0031/954/31954480.ipynb_qa_3"
+kaggle_dataset_name = "uciml/iris"
+gold_answer = "Iris-virginica"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "uciml__iris"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "uciml/iris"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "Iris-virginica"
+QUESTION = "Which species exhibits the highest average petal width according to the feature averages 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/0033_183_33183014_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):
+- WA_Fn-UseC_-Telco-Customer-Churn.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the average tenure duration for customers with a Two-year contract?
+
+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/0033_183_33183014_qa_4/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0033_183_33183014_qa_4"
+description = "What is the average tenure duration for customers with a Two-year contract?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0033/183/33183014.ipynb_qa_4"
+kaggle_dataset_name = "blastchar/telco-customer-churn"
+gold_answer = "56.74"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "blastchar__telco-customer-churn"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "blastchar/telco-customer-churn"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "56.74"
+QUESTION = "What is the average tenure duration for customers with a Two-year contract?"
+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/0038_417_38417457_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):
+- kc_house_data.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which feature has the highest F-value in the SelectKBest feature selection analysis for predicting house prices?
+
+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/0038_417_38417457_qa_1/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0038_417_38417457_qa_1"
+description = "Which feature has the highest F-value in the SelectKBest feature selection analysis for predicting house prices?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0038/417/38417457.ipynb_qa_1"
+kaggle_dataset_name = "harlfoxem/housesalesprediction"
+gold_answer = "sqft_living"
+reward_mode_initial = "exact_short"
+package_tier = 2
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "harlfoxem__housesalesprediction"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "harlfoxem/housesalesprediction"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "sqft_living"
+QUESTION = "Which feature has the highest F-value in the SelectKBest feature selection analysis for predicting house prices?"
+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/0039_815_39815076_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):
+- data.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which algorithm parameter ('auto', 'ball_tree', 'kd_tree', or 'brute') yields the highest model accuracy for the KNN classifier according to the experimental results?
+
+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/0039_815_39815076_qa_3/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0039_815_39815076_qa_3"
+description = "Which algorithm parameter ('auto', 'ball_tree', 'kd_tree', or 'brute') yields the highest model accuracy for the KNN classifier according to the experimental results?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0039/815/39815076.ipynb_qa_3"
+kaggle_dataset_name = "uciml/breast-cancer-wisconsin-data"
+gold_answer = "auto"
+reward_mode_initial = "exact_short"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "uciml__breast-cancer-wisconsin-data"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "uciml/breast-cancer-wisconsin-data"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "auto"
+QUESTION = "Which algorithm parameter ('auto', 'ball_tree', 'kd_tree', or 'brute') yields the highest model accuracy for the KNN classifier according to the experimental results?"
+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/0041_323_41323136_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):
+- data.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the highest F1 score achieved by any of the naive classification baselines presented in 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/0041_323_41323136_qa_3/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-train-v1/0041_323_41323136_qa_3"
+description = "What is the highest F1 score achieved by any of the naive classification baselines presented in the analysis?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0041/323/41323136.ipynb_qa_3"
+kaggle_dataset_name = "uciml/breast-cancer-wisconsin-data"
+gold_answer = "0.5417"
+reward_mode_initial = "numeric"
+package_tier = 2
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "uciml__breast-cancer-wisconsin-data"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "uciml/breast-cancer-wisconsin-data"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "0.5417"
+QUESTION = "What is the highest F1 score achieved by any of the naive classification baselines presented in the analysis?"
+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/0044_153_44153291_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):
+- tmdb_5000_movies.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the mean vote average (C value) calculated across all movies 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/0044_153_44153291_qa_4/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify/0044_153_44153291_qa_4"
+description = "What is the mean vote average (C value) calculated across all movies in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0044/153/44153291.ipynb_qa_4"
+kaggle_dataset_name = "tmdb/tmdb-movie-metadata"
+gold_answer = "6.092"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "tmdb__tmdb-movie-metadata"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "tmdb/tmdb-movie-metadata"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "6.092"
+QUESTION = "What is the mean vote average (C value) calculated across all movies 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/0044_860_44860367_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):
+- vgsales.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which publisher holds the largest market share according to the top 10 publishers 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/0044_860_44860367_qa_5/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0044_860_44860367_qa_5"
+description = "Which publisher holds the largest market share according to the top 10 publishers analysis?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0044/860/44860367.ipynb_qa_5"
+kaggle_dataset_name = "gregorut/videogamesales"
+gold_answer = "Nintendo"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "gregorut__videogamesales"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "gregorut/videogamesales"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "Nintendo"
+QUESTION = "Which publisher holds the largest market share according to the top 10 publishers 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/0048_101_48101055_qa_3/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.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which weekdays have the lowest average sales compared to other weekdays?
+
+Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
+
+Answer as a comma-separated list of the weekday 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/0048_101_48101055_qa_3/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "train-verify2/0048_101_48101055_qa_3"
+description = "Which weekdays have the lowest average sales compared to other weekdays?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0048/101/48101055.ipynb_qa_3"
+kaggle_dataset_name = "carrie1/ecommerce-data"
+gold_answer = "Sunday, Monday"
+reward_mode_initial = "list"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "carrie1__ecommerce-data"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "carrie1/ecommerce-data"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "Sunday, Monday"
+QUESTION = "Which weekdays have the lowest average sales compared to other weekdays?"
+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]