Nyandwi/qwen3.5-9b-data-agent-subset100-eval

Qwen3.5-2B on data_agent_rl_environment_train_subset_100

pass@1: 89/100 = 89% · model Qwen/Qwen3.5-9B · temperature 0.7 · 1 rollout/task · max 10 code turns

Task environments ran as Modal sandboxes (one container per task, Kaggle slice pulled from the HF bucket into /home/user/input). Rewards come from each task's own tests/grader.py with the LLM-judge tier disabled, so grading is deterministic: exact string match, else numeric match within 1e-3.

Where the score comes from

	count	share
Solved	89	89%
↳ graded exact	68	68%
↳ graded numeric (tolerance)	21	21%
Failed	11	11%

Failure breakdown

Category	Count	What it means
solved	89	Answer matched gold (exact or within numeric tolerance).
close_miss	7	Right method, wrong final value (<5% relative error) — usually a filter, rounding or dedup difference.
wrong_value	4	Committed to a numerically distant value — wrong column, wrong aggregation, or wrong interpretation of the question.

Did the model commit to an answer on its own?

The agent scaffold adds a fallback: if the model burns all its code turns without submitting, it gets one final turn that asks only for the answer. This matters for reading the raw number — a 2B model often computes the right value and then keeps exploring instead of committing.

Answer came from	count	of which correct
model submitted <answer> itself	98	88
forced final turn	2	1
read from /workdir/answer.txt	0	0

So 88/100 (88%) were solved with no scaffold intervention; the forced turn recovered a further 1.

Sandbox behaviour

• Mean code cells executed per task: 3.6
• Tasks where at least one cell raised: 12/100
• Tasks that re-ran an identical cell (loop symptom): 9/100
• Mean wall time per task: 19s end-to-end, including container start and the data pull (the first task per image pays ~30s; Modal's cache amortises the rest)
• Harness errors: 0

What's actually going wrong

1. Answering before looking. The median rollout runs just 3 code cells. 7 tasks were answered after a single cell, and 0 of those were wrong. Failed rollouts average 3.8 cells vs 3.5 for solved ones — the model either gets it immediately or flails. Typical shape: it picks a plausible-looking column without checking dtypes/head() and aggregates the wrong thing (e.g. reporting the max of the wrong price column, or counting rows after a filter the question never asked for).

2. Precision conventions, not reasoning. 6 of the 7 close_miss failures are the model being more precise than the gold: the gold answers were rounded by the original notebook author, and the grader's 1e-3 tolerance rejects the full-precision value. These are arguably dataset artifacts rather than model errors — worth deciding explicitly before using this reward signal for RL, since it teaches the model to guess a rounding convention.

3. Bailing out. 0 tasks were answered Not Applicable on questions that were perfectly answerable — in one case the answer was simply the row count. A further 0 returned an uninterpolated template placeholder instead of a value.

4. Tool use is mostly fine. 12/100 tasks hit at least one Python exception, and failed rollouts raise more often (0.18 vs 0.12 exceptions), but exceptions are a symptom of flailing rather than the primary cause. Only 9 tasks showed the repeat-the-same-cell loop.

All 100 tasks

R = reward. Answers and questions are truncated for readability; full trajectories are in trajectories/.

R	Task	Question	Gold	Predicted	Category	Why	Cells
✅	0000_473_473331_qa_1	What is the total number of police death cases recorded in the dataset?	22800	22800	solved		6
✅	0000_710_710837_qa_4	What is the highest price listed for any product in the dataset?	2598.0	2598.0	solved		1
✅	0000_901_901543_qa_3	How many columns are present in the dataset after removing the 'Timestamp' column?	26	26	solved		2
✅	0001_068_1068916_qa_5	What is the average number of comments per entry in the dataset?	0.038	0.03831063927994461	solved		3
✅	0001_426_1426219_qa_2	What is the difference between the highest and lowest opening prices in the dataset?	47830	47830	solved		4
✅	0001_426_1426833_qa_1	What is the maximum closing price recorded in the Bitcoin price dataset?	2958.11	2958.11	solved		4
✅	0001_593_1593609_qa_5	How many missing values were present in the 'type2' column before imputation?	384	384	solved		1
✅	0001_668_1668567_qa_3	How many unique cereal manufacturers are represented in the dataset?	7	7	solved		3
✅	0001_672_1672543_qa_1	What is the critical value of the chi-square distribution at 95% confidence level with 12 degr…	21.026	21.0261	solved		3
✅	0001_736_1736876_qa_4	How many numeric columns are present in the dataset before additional processing?	13	13	solved		7
✅	0001_893_1893246_qa_3	What is the range of the health rating metric (difference between maximum and minimum values) …	75.662061	75.662061	solved		5
✅	0001_903_1903160_qa_3	What is the median rating value for cereals in the dataset?	40.400208	40.400208	solved		2
✅	0002_105_2105396_qa_1	How many missing values were present in the 'Item_Weight' column of the training dataset befor…	1463	1463	solved		2
✅	0013_523_13523088_qa_4	What is the interquartile range (IQR) for the SepalWidthCm feature?	0.5	0.5	solved		3
✅	0013_523_13523088_qa_5	What is the correlation coefficient between Sepal Length and Petal Length in the dataset?	0.871754	0.8717541573048718	solved		4
✅	0015_404_15404226_qa_4	What is the highest OvertimePay amount recorded in the dataset?	245131.88	245131.88	solved		5
✅	0019_106_19106342_qa_1	How many distinct years are covered in the global religious population dataset?	14	14	solved		3
✅	0022_928_22928479_qa_2	How many features in the dataset contain missing values according to the null value analysis?	1	1	solved		8
✅	0023_251_23251738_qa_3	What is the maximum observed value of the 'K' chemical composition feature in the glass sample…	6.21	6.21	solved		5
✅	0023_426_23426961_qa_3	How many columns in the training dataset contain missing values that were identified during pr…	1	1	solved		1
✅	0023_721_23721314_qa_1	What is the highest single-season passing yardage achieved by a player in the dataset?	5477.0	`` tag in the response…`	solved		4
✅	0024_028_24028302_qa_2	What is the total number of drug-related deaths recorded in the dataset?	3583	3583	solved		4
✅	0026_546_26546219_qa_5	What is the average number of bedrooms in the houses included in the dataset?	3.37	3.37084162309721	solved		1
✅	0028_150_28150461_qa_3	How many distinct fruit categories are present in the dataset according to the unique values i…	4	4	solved		4
✅	0030_791_30791695_qa_4	What is the maximum quantity ordered in the dataset?	97	97	solved		5
✅	0032_178_32178295_qa_2	What is the maximum price of a diamond in the dataset before removing rows with zero dimension…	18823	18823	solved		3
✅	0032_340_32340135_qa_5	What is the total number of datasets in the original Kaggle dataset collection analyzed in thi…	2150	2150	solved		2
✅	0035_645_35645519_qa_3	What is the mean pixel intensity value of the first pixel (pixel1) across all images in the tr…	145.419377	145.41937716262976	solved		3
✅	0039_357_39357103_qa_3	What is the median value of the 'concave points_worst' feature across all samples in the datas…	0.09993	0.09993	solved		3
✅	0039_470_39470080_qa_2	What is the average base total statistic across all Pokémon in the dataset?	428.38	428.3770287141074	solved		3
✅	0039_470_39470080_qa_3	What is the 75th percentile value for the special attack (sp_attack) attribute among all Pokém…	91	91	solved		1
✅	0039_729_39729863_qa_4	How many non-null entries exist in the "Unnamed: 2" column of the original dataset?	50	50	solved		3
✅	0039_850_39850014_qa_2	How many missing values are present in the 'event_type' column of the dataset?	4819	4819	solved		2
✅	0040_785_40785152_qa_5	What is the minimum value of 'fixed acidity' in the dataset?	4.6	4.6	solved		4
✅	0041_020_41020843_qa_4	How many devices in the dataset have a price_range classification of 3?	500	500	solved		3
✅	0041_159_41159247_qa_2	What is the most common number of cylinders in the original dataset before outlier removal?	4	4	solved		4
✅	0041_600_41600293_qa_3	How many unique categories are present in the PaymentMethod column of the dataset?	4	4	solved		4
✅	0042_488_42488426_qa_2	What is the total number of hourly records in the Beijing dataset before removing missing valu…	52584	52584	solved		4
✅	0044_367_44367279_qa_2	What is the average tenure (in months) of customers in the dataset?	32.37	32.37114865824223	solved		3
✅	0048_651_48651619_qa_4	What is the median age of all patients in the dataset?	52	52	solved		9
✅	0048_917_48917297_qa_3	How many features are included in the model after removing 'animal_name' and 'class_type'?	16	16	solved		4
✅	0049_261_49261135_qa_3	How many features remain in the dataset after removing 'id', 'diagnosis', and 'Unnamed: 32' co…	30	30	solved		2
✅	0052_080_52080377_qa_4	What is the correlation between Speed and Special Attack (Sp. Atk) in the dataset?	0.4730	0.47301788366869923	solved		3
✅	0052_907_52907926_qa_4	How many numerical features are used as input variables for the machine learning models in thi…	4	4	solved		1
✅	0053_426_53426290_qa_5	How many different stock symbols are present in the dataset?	501	501	solved		3
✅	0054_739_54739768_qa_1	How many missing values were present in the original training dataset before the NaN value cle…	1	1	solved		3
✅	0057_632_57632820_qa_1	Which wine quality rating is the most frequent in the dataset?	5	5	solved		2
✅	0060_532_60532216_qa_4	What is the average body mass index (BMI) of patients in the dataset?	30.66	30.66339686098655	solved		3
✅	0062_251_62251582_qa_2	What was the exact number of missing values in the MINIMUM_PAYMENTS column before imputation?	313	313	solved		5
✅	0064_664_64664569_qa_4	What is the total number of unique values present in the 'stalk-root' feature, including the m…	5	5	solved		5
✅	0064_682_64682826_qa_2	What is the population of the most populous district in the dataset?	11060148	11060148	solved		5
✅	0065_074_65074869_qa_4	Which floor value (e.g., 1.0, 2.0, etc.) has the highest frequency in the dataset?	1.0	1.0	solved		5
✅	0065_150_65150655_qa_3	How many unique classes are present in the dataset based on the label distribution?	24	24	solved		3
✅	0066_186_66186009_qa_2	What is the maximum UnitPrice recorded in the dataset?	38970.00	38970.0	solved		2
✅	0067_118_67118977_qa_5	What is the highest frequency of any wine quality rating in the original dataset?	681	681	solved		3
✅	0067_657_67657879_qa_3	What is the mean of the total length (Petal Length + Sepal Length) across all samples?	9.602	9.602	solved		3
✅	0079_574_79574675_qa_4	What is the most common outcome (0=No diabetes, 1=Diabetes) in the original dataset?	0	0	solved		2
✅	0081_701_81701476_qa_5	What is the kurtosis value for the Petal Length feature in the dataset?	-1.401921	-1.4019208006454036	solved		4
✅	0084_727_84727011_qa_2	How many distinct quality categories exist in the wine quality dataset?	6	6	solved		3
✅	0087_644_87644639_qa_1	What is the highest Speed value recorded in the Pokémon dataset?	180	180	solved		2
✅	0087_867_87867765_qa_2	What is the median living area (sqft_living) of the houses in the dataset?	1910	1910.0	solved		2
✅	0093_755_93755212_qa_5	What is the standard deviation of the age distribution in the dataset?	14.049960	14.049960379216154	solved		4
✅	0094_698_94698351_qa_2	What is the median North American sales value in the dataset?	0.08	0.08	solved		4
✅	0097_234_97234597_qa_3	What is the covariance between Sepal Length and Petal Width in the raw (non-standardized) iris…	0.5169	0.5169038031319911	solved		5
✅	0099_460_99460028_qa_2	What is the median global sales value across all games in the dataset?	0.17	0.17	solved		3
✅	0099_607_99607727_qa_5	What is the median North American sales value across all games in the dataset?	0.08	0.08	solved		5
✅	0103_147_103147643_qa_3	What is the highest recorded final grade (G3) achieved by any student in the dataset?	19	19	solved		3
✅	0104_159_104159135_qa_4	What is the average (mean) year of release for video games in this dataset?	2006.406443	2006.4064433147546	solved		2
✅	0104_601_104601398_qa_4	How many missing values were present in the 'open' column before data cleaning?	11	11	solved		2
✅	0105_079_105079385_qa_4	How many unique categorical values existed in the Item_Type column before label encoding?	16	16	solved		3
✅	0105_135_105135715_qa_2	What was the mean value used to fill missing entries in the 'total_bedrooms' column before app…	538	537.8705525375618	solved		9
✅	0107_517_107517636_qa_3	How many video games in the dataset originally contained missing values in the 'Year' column b…	271	271	solved		3
✅	0107_772_107772825_qa_3	How many patients were removed from the dataset due to invalid age entries (age -1 and 115)?	6	6	solved		8
✅	0108_620_108620410_qa_3	What is the median North American sales value for all video games in the dataset?	0.08	0.08	solved		8
✅	0109_303_109303873_qa_5	What is the median speed value across all Pokémon in the dataset?	65	65.0	solved		2
✅	0111_258_111258803_qa_4	How many entries in the dataset have missing values in the 'Year' column?	271	271	solved		3
✅	0112_685_112685295_qa_4	How many video games are categorized as "SuperHit" based on global sales (≥60 million)?	1	1	solved		3
✅	0113_964_113964418_qa_5	What is the average residual sugar content of the wines in the dataset?	2.5388	2.53880550343965	solved		3
✅	0116_621_116621192_qa_5	What is the total number of video game entries recorded in the dataset?	16,598	16598	solved		3
✅	0116_897_116897141_qa_3	What was the number of missing values present in the 'total_bedrooms' column before data clean…	207	207	solved		3
✅	0119_894_119894753_qa_4	How many entries in the dataset contain missing values in the "Year" column?	271	271	solved		3
✅	0123_734_123734284_qa_1	What is the total number of samples in the Boston housing dataset?	506	506	solved		8
✅	0125_527_125527676_qa_3	What is the median value of Sepal Length (cm) across all samples in the dataset?	5.80	5.8	solved		3
✅	0126_841_126841685_qa_4	How many passengers had missing values in the Age column before imputation?	177	177	solved		4
✅	0127_775_127775252_qa_3	What is the maximum global sales value recorded for any single video game in the dataset?	82.74	82.74	solved		5
✅	0131_454_131454088_qa_5	How many customers in the dataset made a purchase?	143	143	solved		3
✅	0132_624_132624870_qa_4	How many entries in the dataset have missing values in the 'Year' column?	271	271	solved		3
✅	0135_926_135926873_qa_2	What is the correlation coefficient between sales and quantity ordered in the dataset?	0.5514261919183568	0.5514261919183568	solved		3
✅	0136_483_136483047_qa_4	What is the total number of missing values across all columns in the dataset?	0	0	solved		1
❌	0000_683_683069_qa_4	How many hospitals in the dataset have proprietary ownership?	796	800	close_miss	800 vs 796 (0.5% off)	3
❌	0001_532_1532619_qa_4	What is the difference in ROC AUC scores between the optimized SVM model (0.577) and the optim…	0.115	-0.115	wrong_value	-0.115 vs 0.115 (200.0% off)	2
❌	0001_604_1604140_qa_1	What is the average number of total victims per mass shooting incident in the dataset?	10	10.1875	close_miss	10.1875 vs 10 (1.9% off)	5
❌	0014_089_14089673_qa_5	What is the total number of bombing missions recorded in the dataset?	178281	21542	wrong_value	21542 vs 178281 (87.9% off)	10
❌	0015_420_15420402_qa_1	What is the correlation coefficient between Glucose levels and Outcome in the dataset?	0.47	0.46658139830687406	close_miss	0.466581 vs 0.47 (0.7% off)	2
❌	0065_366_65366007_qa_1	What is the mean Heating Load in the building energy efficiency dataset?	22.307201	24.58776041666667	wrong_value	24.5878 vs 22.3072 (10.2% off)	2
❌	0075_081_75081757_qa_3	What is the correlation coefficient between Glucose levels and diabetes diagnosis (Outcome) in…	0.46	0.46658139830687406	close_miss	0.466581 vs 0.46 (1.4% off)	3
❌	0075_568_75568531_qa_3	What is the average number of open accounts held by customers in the dataset?	11.14	11.12853	close_miss	11.1285 vs 11.14 (0.1% off)	3
❌	0094_237_94237964_qa_2	What is the correlation coefficient between carat weight and diamond price in the original dat…	0.947	0.9215913011934763	close_miss	0.921591 vs 0.947 (2.7% off)	3
❌	0094_237_94237964_qa_5	What is the skewness value of the y feature (diamond height in mm) in the original dataset?	2.43	2.434166716488555	close_miss	2.43417 vs 2.43 (0.2% off)	4
❌	0121_209_121209389_qa_2	How many unique categories are present in the 'safety' feature before encoding?	3	4	wrong_value	4 vs 3 (33.3% off)	5