Add files using upload-large-folder tool

README.md

@@ -10,7 +10,9 @@ tags:
 
 # data_agent_harbor_train
 
-5000 deterministic data-analysis training tasks for agent RL. Each task gives an agent a Kaggle dataset and a question; the answer is graded deterministically (exact -> numeric tolerance -> symbolic/math-verify, **no LLM judge**).
+5000 deterministic data-analysis tasks for agent RL (training split). Each task gives an agent a Kaggle dataset and a question; the answer is graded deterministically (exact -> numeric tolerance -> list/percent normalization -> symbolic, **no LLM judge**).
+
+Difficulty tiers: {'easy': 3689, 'hard': 722, 'medium': 589}. Environments build from base image `savatar101/env-data-agent-train:base`.
 
 ## Format
-Harbor task suite: `tasks/<id>/`, `registry.json`, `manifest.parquet`. Environments build from base image `savatar101/env-data-agent-train:base`.
+Harbor task suite: `tasks/<id>/` (task.toml, instruction.md, environment/, tests/), `registry.json`, `manifest.parquet`.

tasks/0001_435_1435960_qa_3/task.toml

@@ -15,6 +15,7 @@ gold_answer = "Never-worked"
 reward_mode_initial = "exact_short"
 package_tier = 2
 difficulty_level = 2
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

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/0010_191_10191909_qa_1/task.toml

@@ -15,6 +15,7 @@ gold_answer = "62.7417"
 reward_mode_initial = "numeric"
 package_tier = 1
 difficulty_level = 2
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0012_593_12593899_qa_2/task.toml

@@ -15,6 +15,7 @@ gold_answer = "1"
 reward_mode_initial = "numeric"
 package_tier = 1
 difficulty_level = 2
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0016_971_16971389_qa_1/task.toml

@@ -15,6 +15,7 @@ gold_answer = "odor"
 reward_mode_initial = "exact_short"
 package_tier = 1
 difficulty_level = 3
+difficulty_tier = "medium"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0025_564_25564899_qa_4/task.toml

@@ -15,6 +15,7 @@ gold_answer = "Logistic Regression"
 reward_mode_initial = "exact_short"
 package_tier = 1
 difficulty_level = 4
+difficulty_tier = "hard"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0026_887_26887570_qa_4/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):
+- Train.csv
+- Test.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+Which outlet size category has the highest median Item_Outlet_Sales according to the pivot table 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/0026_887_26887570_qa_4/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-eval-v1/0026_887_26887570_qa_4"
+description = "Which outlet size category has the highest median Item_Outlet_Sales according to the pivot table analysis?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0026/887/26887570.ipynb_qa_4"
+kaggle_dataset_name = "brijbhushannanda1979/bigmart-sales-data"
+gold_answer = "Medium"
+reward_mode_initial = "exact_short"
+package_tier = 1
+difficulty_level = 2
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "brijbhushannanda1979__bigmart-sales-data"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "brijbhushannanda1979/bigmart-sales-data"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "Medium"
+QUESTION = "Which outlet size category has the highest median Item_Outlet_Sales according to the pivot table 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/0027_179_27179583_qa_2/task.toml

@@ -15,6 +15,7 @@ gold_answer = "28.68"
 reward_mode_initial = "numeric"
 package_tier = 3
 difficulty_level = 2
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0029_184_29184728_qa_1/task.toml

@@ -15,6 +15,7 @@ gold_answer = "Iris-setosa"
 reward_mode_initial = "exact_short"
 package_tier = 1
 difficulty_level = 2
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0029_630_29630344_qa_4/task.toml

@@ -15,6 +15,7 @@ gold_answer = "5268"
 reward_mode_initial = "numeric"
 package_tier = 1
 difficulty_level = 2
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0038_189_38189340_qa_2/task.toml

@@ -15,6 +15,7 @@ gold_answer = "0.04"
 reward_mode_initial = "numeric"
 package_tier = 1
 difficulty_level = 4
+difficulty_tier = "hard"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0043_313_43313326_qa_1/task.toml

@@ -15,6 +15,7 @@ gold_answer = "BMI"
 reward_mode_initial = "exact_short"
 package_tier = 2
 difficulty_level = 3
+difficulty_tier = "medium"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0044_153_44153291_qa_4/task.toml

@@ -15,6 +15,7 @@ gold_answer = "6.092"
 reward_mode_initial = "numeric"
 package_tier = 1
 difficulty_level = 1
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0044_367_44367279_qa_5/task.toml

@@ -15,6 +15,7 @@ gold_answer = "11"
 reward_mode_initial = "numeric"
 package_tier = 2
 difficulty_level = 2
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0044_860_44860367_qa_5/task.toml

@@ -15,6 +15,7 @@ gold_answer = "Nintendo"
 reward_mode_initial = "exact_short"
 package_tier = 1
 difficulty_level = 2
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0047_266_47266614_qa_3/task.toml

@@ -15,6 +15,7 @@ gold_answer = "14"
 reward_mode_initial = "numeric"
 package_tier = 1
 difficulty_level = 4
+difficulty_tier = "hard"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0049_008_49008236_qa_5/task.toml

@@ -15,6 +15,7 @@ gold_answer = "object"
 reward_mode_initial = "exact_short"
 package_tier = 1
 difficulty_level = 1
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0050_784_50784361_qa_5/task.toml

@@ -15,6 +15,7 @@ gold_answer = "77.01"
 reward_mode_initial = "numeric"
 package_tier = 1
 difficulty_level = 1
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0052_980_52980235_qa_2/instruction.md

@@ -11,6 +11,8 @@ What are the four features selected by Recursive Feature Elimination (RFE) when
 
 Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
 
+Answer as a comma-separated list of the exact column 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/0052_980_52980235_qa_2/task.toml

@@ -15,6 +15,7 @@ gold_answer = "battery_power, px_height, px_width, ram"
 reward_mode_initial = "list"
 package_tier = 1
 difficulty_level = 4
+difficulty_tier = "hard"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0055_474_55474406_qa_4/task.toml

@@ -15,6 +15,7 @@ gold_answer = "18"
 reward_mode_initial = "numeric"
 package_tier = 1
 difficulty_level = 2
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0057_712_57712524_qa_5/task.toml

@@ -15,6 +15,7 @@ gold_answer = "5"
 reward_mode_initial = "exact_short"
 package_tier = 1
 difficulty_level = 1
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0057_915_57915748_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):
+- train.csv
+- test.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the interquartile range (IQR) for the 'fc' feature in the original training 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/0057_915_57915748_qa_2/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-eval-v1/0057_915_57915748_qa_2"
+description = "What is the interquartile range (IQR) for the 'fc' feature in the original training dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0057/915/57915748.ipynb_qa_2"
+kaggle_dataset_name = "iabhishekofficial/mobile-price-classification"
+gold_answer = "6.0"
+reward_mode_initial = "numeric"
+package_tier = 2
+difficulty_level = 2
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "iabhishekofficial__mobile-price-classification"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "iabhishekofficial/mobile-price-classification"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "6.0"
+QUESTION = "What is the interquartile range (IQR) for the 'fc' feature in the original training 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/0061_020_61020399_qa_5/task.toml

@@ -15,6 +15,7 @@ gold_answer = "1998"
 reward_mode_initial = "numeric"
 package_tier = 1
 difficulty_level = 1
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0061_164_61164683_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):
+- species.csv
+- parks.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the maximum biodiversity count (number of species) recorded in any national park 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/0061_164_61164683_qa_1/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-eval-v1/0061_164_61164683_qa_1"
+description = "What is the maximum biodiversity count (number of species) recorded in any national park in the dataset?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0061/164/61164683.ipynb_qa_1"
+kaggle_dataset_name = "nationalparkservice/park-biodiversity"
+gold_answer = "6623"
+reward_mode_initial = "numeric"
+package_tier = 1
+difficulty_level = 2
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "nationalparkservice__park-biodiversity"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "nationalparkservice/park-biodiversity"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "6623"
+QUESTION = "What is the maximum biodiversity count (number of species) recorded in any national park 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/0065_989_65989824_qa_2/task.toml

@@ -15,6 +15,7 @@ gold_answer = "34"
 reward_mode_initial = "numeric"
 package_tier = 1
 difficulty_level = 1
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0068_463_68463534_qa_5/task.toml

@@ -15,6 +15,7 @@ gold_answer = "29.0664"
 reward_mode_initial = "numeric"
 package_tier = 1
 difficulty_level = 3
+difficulty_tier = "medium"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0071_619_71619618_qa_1/task.toml

@@ -15,6 +15,7 @@ gold_answer = "86.4"
 reward_mode_initial = "numeric"
 package_tier = 1
 difficulty_level = 2
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0072_665_72665847_qa_1/task.toml

@@ -15,6 +15,7 @@ gold_answer = "below 25 and above 60"
 reward_mode_initial = "exact_short"
 package_tier = 2
 difficulty_level = 3
+difficulty_tier = "medium"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0074_849_74849685_qa_5/task.toml

@@ -15,6 +15,7 @@ gold_answer = "0.90"
 reward_mode_initial = "numeric"
 package_tier = 2
 difficulty_level = 4
+difficulty_tier = "hard"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0077_902_77902994_qa_2/task.toml

@@ -15,6 +15,7 @@ gold_answer = "1.0029"
 reward_mode_initial = "numeric"
 package_tier = 1
 difficulty_level = 1
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0085_729_85729561_qa_4/task.toml

@@ -15,6 +15,7 @@ gold_answer = "Wii Sports"
 reward_mode_initial = "exact_short"
 package_tier = 0
 difficulty_level = 2
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0097_932_97932913_qa_1/instruction.md

@@ -10,6 +10,8 @@ Which two features in the dataset have the highest positive correlation with eac
 
 Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
 
+Answer as a comma-separated list: <feature1>, <feature2>, <correlation> (exact column names, correlation as a plain number).
+
 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/0097_932_97932913_qa_1/task.toml

@@ -15,6 +15,7 @@ gold_answer = "sqft_living, sqft_above, 0.88"
 reward_mode_initial = "list"
 package_tier = 1
 difficulty_level = 3
+difficulty_tier = "medium"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0106_414_106414508_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):
+- car_evaluation.csv
+
+Installed: pandas, numpy, matplotlib, seaborn, scipy, scikit-learn, statsmodels, tabulate, sqlite3, plotly (pip install more if needed).
+
+Question:
+What is the minimum value of the 'buyPrice' feature after ordinal encoding and MinMax scaling?
+
+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/0106_414_106414508_qa_5/task.toml

@@ -0,0 +1,64 @@
+schema_version = "1.2"
+artifacts = []
+
+[task]
+name = "data-agent-eval-v1/0106_414_106414508_qa_5"
+description = "What is the minimum value of the 'buyPrice' feature after ordinal encoding and MinMax scaling?"
+authors = []
+keywords = ["data-agent", "data-analysis", "kaggle"]
+
+[metadata]
+source_dataset = "jupyter-agent/jupyter-agent-dataset"
+source_row_id = "0106/414/106414508.ipynb_qa_5"
+kaggle_dataset_name = "elikplim/car-evaluation-data-set"
+gold_answer = "0.0"
+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 \"$(ls /home/user/input)\" ]"
+interval_sec = 2.0
+timeout_sec = 180.0
+start_period_sec = 5.0
+start_interval_sec = 2.0
+retries = 30
+
+[environment.env]
+HF_BUCKET = "AdithyaSK/jupyter-agent-kaggle-all"
+BUCKET_PREFIX = "elikplim__car-evaluation-data-set"
+HF_TOKEN = "${HF_TOKEN}"
+KAGGLE_DATASET_NAME = "elikplim/car-evaluation-data-set"
+
+[verifier]
+timeout_sec = 120.0
+
+[verifier.env]
+EXPECTED_ANSWER = "0.0"
+QUESTION = "What is the minimum value of the 'buyPrice' feature after ordinal encoding and MinMax scaling?"
+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/0109_422_109422676_qa_4/task.toml

@@ -15,6 +15,7 @@ gold_answer = "G"
 reward_mode_initial = "exact_short"
 package_tier = 1
 difficulty_level = 1
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0112_685_112685295_qa_2/task.toml

@@ -15,6 +15,7 @@ gold_answer = "271"
 reward_mode_initial = "numeric"
 package_tier = 0
 difficulty_level = 1
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0114_582_114582781_qa_5/task.toml

@@ -15,6 +15,7 @@ gold_answer = "smoker"
 reward_mode_initial = "exact_short"
 package_tier = 1
 difficulty_level = 4
+difficulty_tier = "hard"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0115_420_115420161_qa_3/instruction.md

@@ -10,6 +10,8 @@ Which platform has the highest cumulative global sales, and what is the total va
 
 Work it out step by step — inspect the data first (head, shape, dtypes), then compute.
 
+Answer as: <platform>, <value> (comma-separated, label first, keep 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/0115_420_115420161_qa_3/task.toml

@@ -15,6 +15,7 @@ gold_answer = "PS2, 1255.64"
 reward_mode_initial = "list"
 package_tier = 3
 difficulty_level = 2
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0

tasks/0125_030_125030755_qa_2/task.toml

@@ -15,6 +15,7 @@ gold_answer = "Psychic"
 reward_mode_initial = "exact_short"
 package_tier = 1
 difficulty_level = 2
+difficulty_tier = "easy"
 
 [environment]
 build_timeout_sec = 600.0