rubric_id stringlengths 12 12 | text stringlengths 1.1k 2.66k | title stringlengths 46 161 | retrieval_text stringlengths 46 161 | source_instance_id stringclasses 21
values | source_repo stringclasses 21
values | source_arm stringclasses 10
values | source_task_sha stringclasses 21
values | source_dataset stringclasses 1
value | library stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|---|
ae330fd07639 | ### Frozen ImageNet backbone applied to tiny (32×32) images without upscaling or fine-tuning
- **Applies when**: `code` -- the pipeline loads a pretrained CNN (MobileNetV2/ResNet/EfficientNet/VGG) with `include_top=False`, sets `trainable = False`, and feeds images at their native small resolution (≤64 px) on a dataset... | Frozen ImageNet backbone applied to tiny (32×32) images without upscaling or fine-tuning | Frozen ImageNet backbone applied to tiny (32×32) images without upscaling or fine-tuning | aerial-cactus-identification | aerial-cactus-identification | h2-mle-interp | 367d554dc65696f2 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
eb4771193a5a | ### Input normalization does not match the pretrained backbone's expected preprocessing
- **Applies when**: `code` -- images are scaled with a hand-written rule (e.g. `np.array(img)/255.0`) and then fed to a Keras/torchvision pretrained network that ships its own `preprocess_input`/normalization constants.
- **Pattern*... | Input normalization does not match the pretrained backbone's expected preprocessing | Input normalization does not match the pretrained backbone's expected preprocessing | aerial-cactus-identification | aerial-cactus-identification | h2-mle-interp | 367d554dc65696f2 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
ee3f98427dc8 | ### Image loading silently swallows failures, allowing image/label misalignment and dropped test rows
- **Applies when**: `code` -- a loop reads image files and wraps the read in `try: ... except: pass` (or `continue`) while labels/ids are taken from a separate array in the original order.
- **Pattern**: `load_images` ... | Image loading silently swallows failures, allowing image/label misalignment and dropped test rows | Image loading silently swallows failures, allowing image/label misalignment and dropped test rows | aerial-cactus-identification | aerial-cactus-identification | h2-mle-interp | 367d554dc65696f2 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
b8c8c4a4e793 | ### Submission ids built from a directory listing instead of the provided sample submission, with no format validation
- **Applies when**: `task` provides `sample_submission.csv` listing every test id, and `code` constructs the submission id column from `os.listdir(test_dir)` or a glob.
- **Pattern**: The script builds... | Submission ids built from a directory listing instead of the provided sample submission, with no format validation | Submission ids built from a directory listing instead of the provided sample submission, with no format validation | aerial-cactus-identification | aerial-cactus-identification | h2-mle-interp | 367d554dc65696f2 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
ef06eefb9c38 | ### Test-time augmentation and epoch count chosen without validation evidence, while no augmentation or checkpointing is used during training
- **Applies when**: `code` -- the fit call uses a fixed `epochs=N` with no `EarlyStopping`/`ModelCheckpoint`/`ReduceLROnPlateau`, no training-time augmentation, and then averages... | Test-time augmentation and epoch count chosen without validation evidence, while no augmentation or checkpointing is used during training | Test-time augmentation and epoch count chosen without validation evidence, while no augmentation or checkpointing is used during training | aerial-cactus-identification | aerial-cactus-identification | h2-mle-interp | 367d554dc65696f2 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
80fb827037c2 | ### Ignoring the image modality entirely in an image-classification competition
- **Applies when**: `task` describes prediction from images (e.g., dermoscopic JPEGs/DICOMs) and provides a metadata CSV alongside the image folders
- **Pattern**: The pipeline reads only `train.csv`/`test.csv` and builds a model from a han... | Ignoring the image modality entirely in an image-classification competition | Ignoring the image modality entirely in an image-classification competition | siim-isic-melanoma-classification | siim-isic-melanoma-classification | h2-mle-interp | 8e872dbe94c53fd3 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
009bc2f6eecf | ### Target-aggregation features keyed on an ID that does not overlap between train and test
- **Applies when**: `code` builds features by grouping the training label (`groupby(id)['target'].agg(...)`) and then maps them onto the test set
- **Pattern**: Patient-level malignancy count/rate is computed from the training l... | Target-aggregation features keyed on an ID that does not overlap between train and test | Target-aggregation features keyed on an ID that does not overlap between train and test | siim-isic-melanoma-classification | siim-isic-melanoma-classification | h2-mle-interp | 8e872dbe94c53fd3 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
19c46bcf9912 | ### No validation split, cross-validation, or reported metric anywhere in the pipeline
- **Applies when**: `code` trains models and writes a submission
- **Pattern**: Every model is fit on the full training set and predicted straight onto test; there is no `train_test_split`, no `StratifiedKFold`/`GroupKFold`, and no `... | No validation split, cross-validation, or reported metric anywhere in the pipeline | No validation split, cross-validation, or reported metric anywhere in the pipeline | siim-isic-melanoma-classification | siim-isic-melanoma-classification | h2-mle-interp | 8e872dbe94c53fd3 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
296357e9dd24 | ### Blending outputs on incompatible scales and calibrations into one ranking
- **Applies when**: `code` combines predictions from multiple estimator types via mean/median/rank-free averaging
- **Pattern**: Raw `Ridge.predict` values clipped to [0,1] (a regression output on an ~1.8%-positive target, so almost all value... | Blending outputs on incompatible scales and calibrations into one ranking | Blending outputs on incompatible scales and calibrations into one ranking | siim-isic-melanoma-classification | siim-isic-melanoma-classification | h2-mle-interp | 8e872dbe94c53fd3 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
db0a89e808c1 | ### Inconsistent / information-destroying encoding and imputation applied separately to train and test
- **Applies when**: `code` defines a feature function that is called independently on the train frame and the test frame
- **Pattern**: `age_approx` is imputed with each frame's *own* median (train median for train, t... | Inconsistent / information-destroying encoding and imputation applied separately to train and test | Inconsistent / information-destroying encoding and imputation applied separately to train and test | siim-isic-melanoma-classification | siim-isic-melanoma-classification | h2-mle-interp | 8e872dbe94c53fd3 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
c72a9acc7c55 | ### Uses hand-crafted global color statistics with a shallow tabular model instead of a CNN on an image-classification competition
- **Applies when**: `task` describes classifying raw images (here 96x96 RGB pathology patches, ~220k train images, AUC metric) **and** `code` builds a feature vector of summary statistics f... | Uses hand-crafted global color statistics with a shallow tabular model instead of a CNN on an image-classification competition | Uses hand-crafted global color statistics with a shallow tabular model instead of a CNN on an image-classification competition | histopathologic-cancer-detection | histopathologic-cancer-detection | h2-mle-interp | a42375d5a750d8ad | mle-bench-lite-LEAKED | mle-rubrics-oracle |
31e08697b51a | ### Feature engineering ignores the competition's explicit label definition and destroys all spatial/texture information
- **Applies when**: `task` states the label depends on a specific sub-region or spatial property (here: "the center 32x32px region ... contains at least one pixel of tumor tissue"), and `code` comput... | Feature engineering ignores the competition's explicit label definition and destroys all spatial/texture information | Feature engineering ignores the competition's explicit label definition and destroys all spatial/texture information | histopathologic-cancer-detection | histopathologic-cancer-detection | h2-mle-interp | a42375d5a750d8ad | mle-bench-lite-LEAKED | mle-rubrics-oracle |
9302b07c2ab4 | ### Final pipeline depends on a pickled artifact produced interactively, so it is not reproducible end-to-end
- **Applies when**: `task` states a designated script (`final.py`) will be executed once at the end, and `code` loads a model/scaler from a local `.pkl`/`.pt` file rather than training within the script
- **Pat... | Final pipeline depends on a pickled artifact produced interactively, so it is not reproducible end-to-end | Final pipeline depends on a pickled artifact produced interactively, so it is not reproducible end-to-end | histopathologic-cancer-detection | histopathologic-cancer-detection | h2-mle-interp | a42375d5a750d8ad | mle-bench-lite-LEAKED | mle-rubrics-oracle |
1a095a99060b | ### Test IDs are taken from a directory listing instead of the provided sample submission, with no alignment check
- **Applies when**: `task` supplies a `sample_submission.csv` and requires "one row per test id" with exactly matching columns; `code` builds the id list via `os.listdir` on the test folder
- **Pattern**: ... | Test IDs are taken from a directory listing instead of the provided sample submission, with no alignment check | Test IDs are taken from a directory listing instead of the provided sample submission, with no alignment check | histopathologic-cancer-detection | histopathologic-cancer-detection | h2-mle-interp | a42375d5a750d8ad | mle-bench-lite-LEAKED | mle-rubrics-oracle |
9f0690fe9dab | ### Errors and the graded file write are both unguarded: bare `except` substitutes constant features, and the submission is written in place rather than atomically
- **Applies when**: `code` wraps per-sample processing in `try/except` with a hard-coded fallback value, and/or writes predictions directly to the graded pa... | Errors and the graded file write are both unguarded: bare `except` substitutes constant features, and the submission is written in place rather than atomically | Errors and the graded file write are both unguarded: bare `except` substitutes constant features, and the submission is written in place rather than atomically | histopathologic-cancer-detection | histopathologic-cancer-detection | h2-mle-interp | a42375d5a750d8ad | mle-bench-lite-LEAKED | mle-rubrics-oracle |
b9628379e929 | ### Frozen ImageNet backbone with a shallow linear head instead of fine-tuning on the domain images
- **Applies when**: `task` is an image-classification competition on a specialized (medical/satellite/microscopy) image domain with tens of thousands of labeled training images, and `code` loads a pretrained CNN
- **Patt... | Frozen ImageNet backbone with a shallow linear head instead of fine-tuning on the domain images | Frozen ImageNet backbone with a shallow linear head instead of fine-tuning on the domain images | ranzcr-clip-catheter-line-classification | ranzcr-clip-catheter-line-classification | h2-mle-interp | cde916c7fa68553e | mle-bench-lite-LEAKED | mle-rubrics-oracle |
26b495543eeb | ### Training on a small arbitrary subsample of the available labeled data
- **Applies when**: `code` reads the full training table but then restricts fitting to a fixed, hard-coded number of rows (e.g. `min(3500, len(train_df))`) with no stated compute constraint or documented ablation
- **Pattern**: The agent silently... | Training on a small arbitrary subsample of the available labeled data | Training on a small arbitrary subsample of the available labeled data | ranzcr-clip-catheter-line-classification | ranzcr-clip-catheter-line-classification | h2-mle-interp | cde916c7fa68553e | mle-bench-lite-LEAKED | mle-rubrics-oracle |
eeb49be9b094 | ### Applying one backbone's preprocessing to a different backbone that expects a different input scale
- **Applies when**: `code` uses two or more pretrained networks (or one network plus a hand-written normalization) and feeds them from a single shared preprocessing function
- **Pattern**: Images are scaled with `(x/1... | Applying one backbone's preprocessing to a different backbone that expects a different input scale | Applying one backbone's preprocessing to a different backbone that expects a different input scale | ranzcr-clip-catheter-line-classification | ranzcr-clip-catheter-line-classification | h2-mle-interp | cde916c7fa68553e | mle-bench-lite-LEAKED | mle-rubrics-oracle |
b614a2b83829 | ### Using a squared-error regressor with output clipping for a ranking (AUC) metric
- **Applies when**: `task` evaluation is ROC AUC (or another rank-based metric) and `code` fits `Ridge`/`LinearRegression`/`ElasticNet` per label and post-processes predictions with `np.clip(pred, 0, 1)`
- **Pattern**: Ridge on highly i... | Using a squared-error regressor with output clipping for a ranking (AUC) metric | Using a squared-error regressor with output clipping for a ranking (AUC) metric | ranzcr-clip-catheter-line-classification | ranzcr-clip-catheter-line-classification | h2-mle-interp | cde916c7fa68553e | mle-bench-lite-LEAKED | mle-rubrics-oracle |
3e9c560bb464 | ### No held-out validation, so the pipeline's quality is never measured before submission
- **Applies when**: `code` trains any model and writes a submission, and the competition provides labeled training data
- **Pattern**: The script fits on 100% of its (already subsampled) training rows and goes straight to test inf... | No held-out validation, so the pipeline's quality is never measured before submission | No held-out validation, so the pipeline's quality is never measured before submission | ranzcr-clip-catheter-line-classification | ranzcr-clip-catheter-line-classification | h2-mle-interp | cde916c7fa68553e | mle-bench-lite-LEAKED | mle-rubrics-oracle |
f27a855a3fd6 | ### Submission file omits the required ID/key column expected by the leaderboard
- **Applies when**: `code` -- the script builds a `DataFrame`/array of predictions and writes it to the submission path.
- **Pattern**: The attempt constructed `pd.DataFrame({'Insult': y_pred})` only, never carrying over the test row ident... | Submission file omits the required ID/key column expected by the leaderboard | Submission file omits the required ID/key column expected by the leaderboard | detecting-insults-in-social-commentary | detecting-insults-in-social-commentary | h2-mle-interp | b0af0e3739a1a77e | mle-bench-lite-LEAKED | mle-rubrics-oracle |
52d2ab6cca78 | ### No read-back / schema assertion on the written submission before finishing
- **Applies when**: `code` -- any pipeline that ends by writing a submission artifact to a fixed path.
- **Pattern**: The script writes the CSV and terminates with zero verification: it never loads `sample_submission.csv` to compare headers,... | No read-back / schema assertion on the written submission before finishing | No read-back / schema assertion on the written submission before finishing | detecting-insults-in-social-commentary | detecting-insults-in-social-commentary | h2-mle-interp | b0af0e3739a1a77e | mle-bench-lite-LEAKED | mle-rubrics-oracle |
3e426e710b4f | ### Two independent model families are averaged 50/50 without ever scoring the blend
- **Applies when**: `code` -- the final prediction is a combination of two or more separately-built prediction vectors (e.g. `y_pred = (y_pred_ensemble + y_pred_stacking) / 2`).
- **Pattern**: The attempt computes OOF AUCs for the indi... | Two independent model families are averaged 50/50 without ever scoring the blend | Two independent model families are averaged 50/50 without ever scoring the blend | detecting-insults-in-social-commentary | detecting-insults-in-social-commentary | h2-mle-interp | b0af0e3739a1a77e | mle-bench-lite-LEAKED | mle-rubrics-oracle |
bec65aceee01 | ### Vectorizer capacity hard-capped far below what the text task needs
- **Applies when**: `task` is text classification over free-form user comments and `code` uses `TfidfVectorizer`/`CountVectorizer` with an explicit `max_features` cap.
- **Pattern**: Every vectorizer is capped at 4000–5000 features with `min_df=2, m... | Vectorizer capacity hard-capped far below what the text task needs | Vectorizer capacity hard-capped far below what the text task needs | detecting-insults-in-social-commentary | detecting-insults-in-social-commentary | h2-mle-interp | b0af0e3739a1a77e | mle-bench-lite-LEAKED | mle-rubrics-oracle |
282c4f9eeefd | ### Aggressive lowercasing/normalization with no compensating engineered features for the task signal
- **Applies when**: `task` involves detecting abusive/insulting comments and `code` applies a `clean_text` function before vectorization.
- **Pattern**: The cleaner lowercases everything and collapses escape sequences,... | Aggressive lowercasing/normalization with no compensating engineered features for the task signal | Aggressive lowercasing/normalization with no compensating engineered features for the task signal | detecting-insults-in-social-commentary | detecting-insults-in-social-commentary | h2-mle-interp | b0af0e3739a1a77e | mle-bench-lite-LEAKED | mle-rubrics-oracle |
cf1afc62eed0 | ### Hardcoded test-file name assumed instead of discovered from the data directory
- **Applies when**: `code` -- the script reads competition inputs with a literal path such as `/home/data/en_test_2.csv` (or any filename not shown in the competition's file description) and there is no directory listing, glob, or existe... | Hardcoded test-file name assumed instead of discovered from the data directory | Hardcoded test-file name assumed instead of discovered from the data directory | text-normalization-challenge-english-language | text-normalization-challenge-english-language | h2-mle-interp | 412eeab709f31579 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
a1b36f90a57c | ### Final submission written directly to the graded path with no early baseline and no atomic move
- **Applies when**: `code` -- the script's only write of the graded artifact is `to_csv('/home/submission/submission.csv')` executed at the end of a long, all-or-nothing pipeline.
- **Pattern**: Nothing is written to the ... | Final submission written directly to the graded path with no early baseline and no atomic move | Final submission written directly to the graded path with no early baseline and no atomic move | text-normalization-challenge-english-language | text-normalization-challenge-english-language | h2-mle-interp | 412eeab709f31579 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
79544266817c | ### Final pipeline depends on pickles produced by earlier ad-hoc commands, with no regeneration path
- **Applies when**: `code` -- the scoring script loads model/state artifacts (`.pkl`, `.npy`, `.joblib`) from the working directory that are not built inside the same script.
- **Pattern**: All three mappings (`class_ma... | Final pipeline depends on pickles produced by earlier ad-hoc commands, with no regeneration path | Final pipeline depends on pickles produced by earlier ad-hoc commands, with no regeneration path | text-normalization-challenge-english-language | text-normalization-challenge-english-language | h2-mle-interp | 412eeab709f31579 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
cc7e64ed6b65 | ### Submission ids reconstructed from columns without verifying coverage against the sample submission
- **Applies when**: `task` provides `sample_submission.csv` and an `id` convention (`sentence_id_token_id`), and `code` builds the submission frame itself.
- **Pattern**: The script builds `id` via `sentence_id.astype... | Submission ids reconstructed from columns without verifying coverage against the sample submission | Submission ids reconstructed from columns without verifying coverage against the sample submission | text-normalization-challenge-english-language | text-normalization-challenge-english-language | h2-mle-interp | 412eeab709f31579 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
935f0c22e678 | ### Heuristic class assignment applied with a broad catch-all last and never validated on labeled training data
- **Applies when**: `task` hides a label/class column at test time and `code` re-infers it with hand-written regex rules, then uses `(class, token)` keys for dictionary lookup.
- **Pattern**: Rules are applie... | Heuristic class assignment applied with a broad catch-all last and never validated on labeled training data | Heuristic class assignment applied with a broad catch-all last and never validated on labeled training data | text-normalization-challenge-english-language | text-normalization-challenge-english-language | h2-mle-interp | 412eeab709f31579 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
a712b31b86c7 | ### Model family never varied beyond tree ensembles on a small, dense, high-dimensional multi-class log-loss task
- **Applies when**: `task` describes a tabular dataset with few hundred–few thousand rows, ~100+ continuous engineered features, many classes, and a probabilistic metric (multi-class log loss); `code` fits ... | Model family never varied beyond tree ensembles on a small, dense, high-dimensional multi-class log-loss task | Model family never varied beyond tree ensembles on a small, dense, high-dimensional multi-class log-loss task | leaf-classification | leaf-classification | h2-mle-interp | 9a6e9ba8c81700ef | mle-bench-lite-LEAKED | mle-rubrics-oracle |
974706550734 | ### Raw tree-ensemble `predict_proba` submitted for log loss with no calibration, smoothing, or clipping
- **Applies when**: `task` is scored by log loss (with probabilities clipped at ~1e-15 by the grader) and `code` writes `model.predict_proba(...)` (or an average of such) straight to the submission.
- **Pattern**: T... | Raw tree-ensemble `predict_proba` submitted for log loss with no calibration, smoothing, or clipping | Raw tree-ensemble `predict_proba` submitted for log loss with no calibration, smoothing, or clipping | leaf-classification | leaf-classification | h2-mle-interp | 9a6e9ba8c81700ef | mle-bench-lite-LEAKED | mle-rubrics-oracle |
98d2597b5ff7 | ### No feature scaling or dimensionality reduction, silently foreclosing the model families that fit this data
- **Applies when**: `code` loads a feature matrix of continuous attributes on differing scales (here 64 margin + 64 shape + 64 texture histogram/descriptor values) and feeds `train.drop(['id','species'])` dire... | No feature scaling or dimensionality reduction, silently foreclosing the model families that fit this data | No feature scaling or dimensionality reduction, silently foreclosing the model families that fit this data | leaf-classification | leaf-classification | h2-mle-interp | 9a6e9ba8c81700ef | mle-bench-lite-LEAKED | mle-rubrics-oracle |
797fbd5b31ba | ### Supplied raw images and the description's explicit feature-engineering hint were ignored
- **Applies when**: `task` ships an auxiliary data modality (an `images/` directory keyed by `id`) and the description explicitly suggests going beyond the pre-extracted features; `code` reads only the CSVs.
- **Pattern**: The ... | Supplied raw images and the description's explicit feature-engineering hint were ignored | Supplied raw images and the description's explicit feature-engineering hint were ignored | leaf-classification | leaf-classification | h2-mle-interp | 9a6e9ba8c81700ef | mle-bench-lite-LEAKED | mle-rubrics-oracle |
48ac8b37097a | ### Final submission script computes no validation and hard-codes a stale CV number, so no "only replace if better" guard exists
- **Applies when**: `instructions` require producing an early baseline and replacing the submission only on validated improvement; `code` is the final script that writes `/home/submission/sub... | Final submission script computes no validation and hard-codes a stale CV number, so no "only replace if better" guard exists | Final submission script computes no validation and hard-codes a stale CV number, so no "only replace if better" guard exists | leaf-classification | leaf-classification | h2-mle-interp | 9a6e9ba8c81700ef | mle-bench-lite-LEAKED | mle-rubrics-oracle |
ac70de301b3a | ### Collapsing a spectrogram into a handful of global scalars for a time–frequency detection task
- **Applies when**: `task` involves detecting a localized acoustic/visual event (e.g., a whale upcall in a 2-second clip, AUC-scored) and `code` builds a tabular feature vector from audio/image files
- **Pattern**: The att... | Collapsing a spectrogram into a handful of global scalars for a time–frequency detection task | Collapsing a spectrogram into a handful of global scalars for a time–frequency detection task | the-icml-2013-whale-challenge-right-whale-redux | the-icml-2013-whale-challenge-right-whale-redux | h2-mle-interp | 6c5760dd0ab95353 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
a33e7788476e | ### No held-out validation, so model/ensemble choices are unverified guesses
- **Applies when**: `code` fits more than one model, blends predictions, or sets hyperparameters, and the task instructions require replacing the submission only on validation improvement
- **Pattern**: The script trains GradientBoosting, Rand... | No held-out validation, so model/ensemble choices are unverified guesses | No held-out validation, so model/ensemble choices are unverified guesses | the-icml-2013-whale-challenge-right-whale-redux | the-icml-2013-whale-challenge-right-whale-redux | h2-mle-interp | 6c5760dd0ab95353 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
4e2eb4c59f2c | ### Broad `try/except` that silently substitutes all-zero feature vectors
- **Applies when**: `code` parses many files (audio, images, text) inside a `try`/`except` block and returns a default value on failure
- **Pattern**: `extract_audio_features` wraps the whole body in `try: ... except Exception: return {f: 0 for f... | Broad `try/except` that silently substitutes all-zero feature vectors | Broad `try/except` that silently substitutes all-zero feature vectors | the-icml-2013-whale-challenge-right-whale-redux | the-icml-2013-whale-challenge-right-whale-redux | h2-mle-interp | 6c5760dd0ab95353 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
e047cb4f815a | ### Final script depends on hand-made working-directory artifacts and writes the submission non-atomically
- **Applies when**: `task` states the data lives at a read-only path (e.g., `/home/data`) and that `final.py` will be re-executed standalone at the end of the run
- **Pattern**: The pipeline hard-codes `/home/agen... | Final script depends on hand-made working-directory artifacts and writes the submission non-atomically | Final script depends on hand-made working-directory artifacts and writes the submission non-atomically | the-icml-2013-whale-challenge-right-whale-redux | the-icml-2013-whale-challenge-right-whale-redux | h2-mle-interp | 6c5760dd0ab95353 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
c677cb2bd8a8 | ### Submission rows assembled from a directory listing rather than aligned to the sample submission
- **Applies when**: `task` supplies a `sample_submission.csv` with an explicit id column and required row set
- **Pattern**: The `clip` column is built from `sorted(os.listdir(test_dir))` filtered by `.aif`, and the file... | Submission rows assembled from a directory listing rather than aligned to the sample submission | Submission rows assembled from a directory listing rather than aligned to the sample submission | the-icml-2013-whale-challenge-right-whale-redux | the-icml-2013-whale-challenge-right-whale-redux | h2-mle-interp | 6c5760dd0ab95353 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
a8fde425d478 | ### Final script substitutes a randomly initialized classifier head and never trains or loads weights for it
- **Applies when**: `code` -- the pipeline builds a pretrained CNN (e.g. `resnet50(weights=...)`) and then overwrites the classifier (`model.fc = nn.Linear(...)`, `model.classifier = ...`) before inference.
- **... | Final script substitutes a randomly initialized classifier head and never trains or loads weights for it | Final script substitutes a randomly initialized classifier head and never trains or loads weights for it | dogs-vs-cats-redux-kernels-edition | dogs-vs-cats-redux-kernels-edition | h2-mle-interp | 7b1e402fb7088f94 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
65096adb6108 | ### Final pipeline never reads the labeled training data
- **Applies when**: `task` provides a labeled `train/` folder and `code` is the script designated as the full final pipeline.
- **Pattern**: The script only enumerates `/home/data/test`, builds a model, and predicts. No path to the training images, no filename-de... | Final pipeline never reads the labeled training data | Final pipeline never reads the labeled training data | dogs-vs-cats-redux-kernels-edition | dogs-vs-cats-redux-kernels-edition | h2-mle-interp | 7b1e402fb7088f94 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
f94b44030dea | ### No held-out validation score is computed, so the submission is replaced without evidence of improvement
- **Applies when**: `task` specifies a metric (log loss) and instructs replacing the submission only when validation improves; `code` writes `/home/submission/submission.csv`.
- **Pattern**: The pipeline contains... | No held-out validation score is computed, so the submission is replaced without evidence of improvement | No held-out validation score is computed, so the submission is replaced without evidence of improvement | dogs-vs-cats-redux-kernels-edition | dogs-vs-cats-redux-kernels-edition | h2-mle-interp | 7b1e402fb7088f94 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
a328cfcffd30 | ### Pretrained ImageNet model is loaded but its free, task-relevant signal is discarded
- **Applies when**: `task` is a binary image classification whose classes map onto ImageNet categories (cats/dogs), and `code` loads ImageNet-pretrained weights.
- **Pattern**: The script pays the cost of downloading ImageNet weight... | Pretrained ImageNet model is loaded but its free, task-relevant signal is discarded | Pretrained ImageNet model is loaded but its free, task-relevant signal is discarded | dogs-vs-cats-redux-kernels-edition | dogs-vs-cats-redux-kernels-edition | h2-mle-interp | 7b1e402fb7088f94 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
91ee0f17b1cc | ### Log-loss submission written with no probability clipping and no guard against degenerate prediction distributions
- **Applies when**: `task` is scored by log loss and `code` writes raw model probabilities to the submission.
- **Pattern**: Probabilities go straight from `sigmoid`/`softmax` into `to_csv` with no `np.... | Log-loss submission written with no probability clipping and no guard against degenerate prediction distributions | Log-loss submission written with no probability clipping and no guard against degenerate prediction distributions | dogs-vs-cats-redux-kernels-edition | dogs-vs-cats-redux-kernels-edition | h2-mle-interp | 7b1e402fb7088f94 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
1bb08c70153c | ### Training on an arbitrary tiny subsample of a very large dataset with no scaling or justification
- **Applies when**: `task` states the training file is very large (e.g. ~55M rows) and `code` reads it with a hard-coded row cap
- **Pattern**: The agent calls `pd.read_csv(..., nrows=500000)` (<1% of available data), t... | Training on an arbitrary tiny subsample of a very large dataset with no scaling or justification | Training on an arbitrary tiny subsample of a very large dataset with no scaling or justification | new-york-city-taxi-fare-prediction | new-york-city-taxi-fare-prediction | h2-mle-interp | b276dc75f33a7ba7 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
2e7f68da0fce | ### No held-out validation, so model quality and ensemble weights are never measured
- **Applies when**: `code` fits models and writes a submission
- **Pattern**: There is no train/validation split, no `cross_val_score`, no `eval_set`/early stopping, and no printed RMSE. Three models are blended with hand-picked weight... | No held-out validation, so model quality and ensemble weights are never measured | No held-out validation, so model quality and ensemble weights are never measured | new-york-city-taxi-fare-prediction | new-york-city-taxi-fare-prediction | h2-mle-interp | b276dc75f33a7ba7 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
0d963d74e002 | ### Discarding raw high-signal columns (absolute coordinates, year) in favour of only derived aggregates
- **Applies when**: `task` supplies raw pickup/dropoff coordinates and a datetime spanning multiple years, and `code` builds an explicit `features` list
- **Pattern**: The feature list contains `distance`, `hour`, `... | Discarding raw high-signal columns (absolute coordinates, year) in favour of only derived aggregates | Discarding raw high-signal columns (absolute coordinates, year) in favour of only derived aggregates | new-york-city-taxi-fare-prediction | new-york-city-taxi-fare-prediction | h2-mle-interp | b276dc75f33a7ba7 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
3a7db2b90bc9 | ### Cleaning the training set on target/distance only, ignoring the corrupt-coordinate rows that dominate this dataset
- **Applies when**: `code` contains row-filtering on the training frame for a dataset documented to contain noisy geolocation
- **Pattern**: Filters are applied to `fare_amount` (2.5–150) and to the co... | Cleaning the training set on target/distance only, ignoring the corrupt-coordinate rows that dominate this dataset | Cleaning the training set on target/distance only, ignoring the corrupt-coordinate rows that dominate this dataset | new-york-city-taxi-fare-prediction | new-york-city-taxi-fare-prediction | h2-mle-interp | b276dc75f33a7ba7 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
b3d8e3a1061c | ### Hard-coded post-hoc prediction overrides and clipping chosen without evidence
- **Applies when**: `code` mutates `y_pred` after model inference (clipping, constant substitution, rule-based overwrite)
- **Pattern**: Predictions are clipped to `[2.5, 150]` — the same bounds used to filter training rows — and every te... | Hard-coded post-hoc prediction overrides and clipping chosen without evidence | Hard-coded post-hoc prediction overrides and clipping chosen without evidence | new-york-city-taxi-fare-prediction | new-york-city-taxi-fare-prediction | h2-mle-interp | b276dc75f33a7ba7 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
183268bdc2b0 | ### Relies on a single precomputed supplemental feature file and ignores the richer representations the competition supplies
- **Applies when**: `task` provides multiple feature views (raw WAV audio, spectrograms, per-segment feature tables, bag-of-instance representations) plus one or more convenience "summary" files,... | Relies on a single precomputed supplemental feature file and ignores the richer representations the competition supplies | Relies on a single precomputed supplemental feature file and ignores the richer representations the competition supplies | mlsp-2013-birds | mlsp-2013-birds | h2-mle-interp | 66f5b5a3d618e9f0 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
ef40da4811f1 | ### No validation split or cross-validated score is ever computed, so every modeling choice is unverified
- **Applies when**: `code` fits models and writes a submission, and the training data has labels available (here `fold == 0` rows have real species lists)
- **Pattern**: The script splits strictly into `fold==0` (f... | No validation split or cross-validated score is ever computed, so every modeling choice is unverified | No validation split or cross-validated score is ever computed, so every modeling choice is unverified | mlsp-2013-birds | mlsp-2013-birds | h2-mle-interp | 66f5b5a3d618e9f0 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
c8527cdf196f | ### Uniformly averages many near-duplicate models from the same two or three families and calls it a "mega ensemble"
- **Applies when**: `code` builds an ensemble by looping over a small hyperparameter grid of the same estimator class and averaging all outputs with equal weight
- **Pattern**: 9 GradientBoosting variant... | Uniformly averages many near-duplicate models from the same two or three families and calls it a "mega ensemble" | Uniformly averages many near-duplicate models from the same two or three families and calls it a "mega ensemble" | mlsp-2013-birds | mlsp-2013-birds | h2-mle-interp | 66f5b5a3d618e9f0 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
7be9b8496f12 | ### Produces per-species probabilities without making them comparable across species, although the metric pools all (rec_id, species) rows into one AUC
- **Applies when**: `task` evaluates a single ROC AUC over a flattened Id column that mixes classes (here `Id = rec_id*100 + species`), and `code` trains one independen... | Produces per-species probabilities without making them comparable across species, although the metric pools all (rec_id, species) rows into one AUC | Produces per-species probabilities without making them comparable across species, although the metric pools all (rec_id, species) rows into one AUC | mlsp-2013-birds | mlsp-2013-birds | h2-mle-interp | 66f5b5a3d618e9f0 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
4d9505249d7e | ### Constructs submission Ids from an internal fold file instead of validating against the official sample submission, with no coverage or degenerate-column guards
- **Applies when**: `task` specifies an exact submission Id set/format, and `code` derives the rows to emit from a metadata file it chose itself
- **Pattern... | Constructs submission Ids from an internal fold file instead of validating against the official sample submission, with no coverage or degenerate-column guards | Constructs submission Ids from an internal fold file instead of validating against the official sample submission, with no coverage or degenerate-column guards | mlsp-2013-birds | mlsp-2013-birds | h2-mle-interp | 66f5b5a3d618e9f0 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
68465b05c18f | ### Word-only TF-IDF on noisy user-generated text (no character n-grams)
- **Applies when**: `task` is text classification over raw user comments/posts (misspellings, obfuscated profanity, leetspeak), and `code` builds features with `TfidfVectorizer`/`CountVectorizer`.
- **Pattern**: The pipeline vectorizes only word u... | Word-only TF-IDF on noisy user-generated text (no character n-grams) | Word-only TF-IDF on noisy user-generated text (no character n-grams) | jigsaw-toxic-comment-classification-challenge | jigsaw-toxic-comment-classification-challenge | h2-mle-interp | 9966a3e9cabce45a | mle-bench-lite-LEAKED | mle-rubrics-oracle |
7ac196b3ae7a | ### Final model fitted on a train split only, discarding the held-out fold
- **Applies when**: `code` calls `train_test_split` (or slices a holdout) to compute a validation score and then uses that same fitted estimator to predict the test set.
- **Pattern**: For each of the six labels the script splits off 20% for val... | Final model fitted on a train split only, discarding the held-out fold | Final model fitted on a train split only, discarding the held-out fold | jigsaw-toxic-comment-classification-challenge | jigsaw-toxic-comment-classification-challenge | h2-mle-interp | 9966a3e9cabce45a | mle-bench-lite-LEAKED | mle-rubrics-oracle |
fcdb6c249dab | ### English stop-word removal and aggressive vocabulary pruning delete the discriminative signal
- **Applies when**: `task` is toxicity/sentiment/abuse detection where pronouns, intensifiers, punctuation and casing carry meaning, and `code` sets `stop_words=`, a low `max_features`, or a restrictive `max_df`/`token_patt... | English stop-word removal and aggressive vocabulary pruning delete the discriminative signal | English stop-word removal and aggressive vocabulary pruning delete the discriminative signal | jigsaw-toxic-comment-classification-challenge | jigsaw-toxic-comment-classification-challenge | h2-mle-interp | 9966a3e9cabce45a | mle-bench-lite-LEAKED | mle-rubrics-oracle |
9291b04ef9ea | ### Hyperparameters declared "optimized" with no search or comparison in the code
- **Applies when**: `code` hard-codes model/vectorizer hyperparameters (`C=0.8`, `max_features=40000`, `max_df=0.98`, `class_weight='balanced'`) and comments describe them as tuned or optimized.
- **Pattern**: There is no `GridSearchCV`, ... | Hyperparameters declared "optimized" with no search or comparison in the code | Hyperparameters declared "optimized" with no search or comparison in the code | jigsaw-toxic-comment-classification-challenge | jigsaw-toxic-comment-classification-challenge | h2-mle-interp | 9966a3e9cabce45a | mle-bench-lite-LEAKED | mle-rubrics-oracle |
42e6638634d9 | ### Single-model pipeline with no blending, despite an averaged-AUC metric that rewards diversity
- **Applies when**: `task` is scored by mean column-wise ROC AUC (or any averaged ranking metric) over several correlated labels, and `code` produces test predictions from exactly one estimator family.
- **Pattern**: Six i... | Single-model pipeline with no blending, despite an averaged-AUC metric that rewards diversity | Single-model pipeline with no blending, despite an averaged-AUC metric that rewards diversity | jigsaw-toxic-comment-classification-challenge | jigsaw-toxic-comment-classification-challenge | h2-mle-interp | 9966a3e9cabce45a | mle-bench-lite-LEAKED | mle-rubrics-oracle |
1814d4469987 | ### Deep-learning-required image task solved with global color statistics + tabular tree ensembles
- **Applies when**: `task` is an image-classification competition (raw jpg/png inputs, fine-grained visual categories such as leaf disease symptoms) **and** `code` builds a fixed-length numeric feature vector per image an... | Deep-learning-required image task solved with global color statistics + tabular tree ensembles | Deep-learning-required image task solved with global color statistics + tabular tree ensembles | plant-pathology-2020-fgvc7 | plant-pathology-2020-fgvc7 | h2-mle-interp | 40d3639cfea9ca36 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
9c2497040cc9 | ### No validation split or cross-validation, so hyperparameters and ensemble weights are set blind
- **Applies when**: `code` trains one or more models and produces a submission, and the competition specifies a scoring metric (mean column-wise ROC AUC)
- **Pattern**: Every model is fit on 100% of the training rows and ... | No validation split or cross-validation, so hyperparameters and ensemble weights are set blind | No validation split or cross-validation, so hyperparameters and ensemble weights are set blind | plant-pathology-2020-fgvc7 | plant-pathology-2020-fgvc7 | h2-mle-interp | 40d3639cfea9ca36 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
4f2ce12996f2 | ### Submission column names/order hard-coded from assumption instead of read from sample_submission.csv
- **Applies when**: `code` constructs the submission DataFrame by writing literal column names, and the competition provides a `sample_submission.csv`
- **Pattern**: The script never opens `sample_submission.csv`. Ta... | Submission column names/order hard-coded from assumption instead of read from sample_submission.csv | Submission column names/order hard-coded from assumption instead of read from sample_submission.csv | plant-pathology-2020-fgvc7 | plant-pathology-2020-fgvc7 | h2-mle-interp | 40d3639cfea9ca36 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
0bd500245133 | ### Test rows conditionally dropped by an `os.path.exists` guard, with no reindex back to the full test id list
- **Applies when**: `code` iterates over the test manifest and appends rows to the prediction set only inside a conditional (file-existence check, try/except, size filter)
- **Pattern**: Test features are col... | Test rows conditionally dropped by an `os.path.exists` guard, with no reindex back to the full test id list | Test rows conditionally dropped by an `os.path.exists` guard, with no reindex back to the full test id list | plant-pathology-2020-fgvc7 | plant-pathology-2020-fgvc7 | h2-mle-interp | 40d3639cfea9ca36 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
ffc279d7e8de | ### Bare `except` returning a fallback vector of the wrong dimension, silently corrupting the feature matrix
- **Applies when**: `code` wraps per-sample feature extraction in `try/except` and returns a constant array on failure
- **Pattern**: `extract_multiscale_features` catches every exception and returns `np.zeros(1... | Bare `except` returning a fallback vector of the wrong dimension, silently corrupting the feature matrix | Bare `except` returning a fallback vector of the wrong dimension, silently corrupting the feature matrix | plant-pathology-2020-fgvc7 | plant-pathology-2020-fgvc7 | h2-mle-interp | 40d3639cfea9ca36 | mle-bench-lite-LEAKED | mle-rubrics-oracle |
2eb5468fe327 | ### Submission file built from directory listing instead of `sample_submission.csv`
- **Applies when**: `code` -- the competition supplies a `sample_submission.csv` with a fixed id order and 120 named probability columns, and the script constructs the submission DataFrame itself
- **Pattern**: The agent derived test id... | Submission file built from directory listing instead of `sample_submission.csv` | Submission file built from directory listing instead of `sample_submission.csv` | dog-breed-identification | dog-breed-identification | h2-mle-interp | f58d30288d9032cc | mle-bench-lite-LEAKED | mle-rubrics-oracle |
ef926af26fc9 | ### Final script writes the submission directly to the graded path with no directory creation, atomic move, or fallback
- **Applies when**: `code` -- the run instructions state that whatever exists at `/home/submission/submission.csv` is graded as-is and that the final script is executed once unattended
- **Pattern**: ... | Final script writes the submission directly to the graded path with no directory creation, atomic move, or fallback | Final script writes the submission directly to the graded path with no directory creation, atomic move, or fallback | dog-breed-identification | dog-breed-identification | h2-mle-interp | f58d30288d9032cc | mle-bench-lite-LEAKED | mle-rubrics-oracle |
e885e3966bdc | ### Unguarded dependency on downloading pretrained weights at run time
- **Applies when**: `code` -- the pipeline calls `models.<arch>(pretrained=True)` / `weights=...`, `timm.create_model(..., pretrained=True)`, or `from_pretrained(...)` inside the final script
- **Pattern**: `models.mobilenet_v2(pretrained=True)` is ... | Unguarded dependency on downloading pretrained weights at run time | Unguarded dependency on downloading pretrained weights at run time | dog-breed-identification | dog-breed-identification | h2-mle-interp | f58d30288d9032cc | mle-bench-lite-LEAKED | mle-rubrics-oracle |
88177695733e | ### No held-out validation, so the pipeline's quality is never measured or compared
- **Applies when**: `task` -- the metric is multi-class log loss and the instructions say to replace the submission only when your own validation score improves
- **Pattern**: The script trains on 100% of `labels.csv` and reports only r... | No held-out validation, so the pipeline's quality is never measured or compared | No held-out validation, so the pipeline's quality is never measured or compared | dog-breed-identification | dog-breed-identification | h2-mle-interp | f58d30288d9032cc | mle-bench-lite-LEAKED | mle-rubrics-oracle |
982270e04b2f | ### Transfer-learning capacity thrown away: tiny input resolution and a single epoch on a fine-grained 120-class task
- **Applies when**: `task` -- fine-grained image classification with ~10k training images, 120 classes, scored by log loss, with a ~2-hour compute budget and an unlimited-time final script
- **Pattern**... | Transfer-learning capacity thrown away: tiny input resolution and a single epoch on a fine-grained 120-class task | Transfer-learning capacity thrown away: tiny input resolution and a single epoch on a fine-grained 120-class task | dog-breed-identification | dog-breed-identification | h2-mle-interp | f58d30288d9032cc | mle-bench-lite-LEAKED | mle-rubrics-oracle |
2fc5739ee7ea | ### High-cardinality string feature collapsed into a single ordinal code instead of being decomposed into its components
- **Applies when**: `code` -- the dataset contains an object/string column (here `f_27`, a fixed-length alphabetic string) and the code converts it with `OrdinalEncoder`/`LabelEncoder`/`astype('categ... | High-cardinality string feature collapsed into a single ordinal code instead of being decomposed into its components | High-cardinality string feature collapsed into a single ordinal code instead of being decomposed into its components | tabular-playground-series-may-2022 | tabular-playground-series-may-2022 | h2-mle-interp | 0ce5bc4eb40895cb | mle-bench-lite-LEAKED | mle-rubrics-oracle |
65612c86a063 | ### Categorical encoder fitted on train only, so unseen test categories all collapse to one sentinel value
- **Applies when**: `code` -- an encoder is `fit` on the training split (or training folds) and `transform`ed on test with `handle_unknown='use_encoded_value', unknown_value=-1` (or `errors='ignore'`, `.map(dict)`... | Categorical encoder fitted on train only, so unseen test categories all collapse to one sentinel value | Categorical encoder fitted on train only, so unseen test categories all collapse to one sentinel value | tabular-playground-series-may-2022 | tabular-playground-series-may-2022 | h2-mle-interp | 0ce5bc4eb40895cb | mle-bench-lite-LEAKED | mle-rubrics-oracle |
0664aad87c3f | ### Interaction features hand-guessed from column names rather than selected from data evidence
- **Applies when**: `task` mentions that the data contains important feature interactions, and `code` builds a hard-coded list of feature pairs/products.
- **Pattern**: The attempt wrote a literal list of ~23 `(f_a, f_b)` pa... | Interaction features hand-guessed from column names rather than selected from data evidence | Interaction features hand-guessed from column names rather than selected from data evidence | tabular-playground-series-may-2022 | tabular-playground-series-may-2022 | h2-mle-interp | 0ce5bc4eb40895cb | mle-bench-lite-LEAKED | mle-rubrics-oracle |
fb5e93aff758 | ### Ensemble weights and model choice fixed without any out-of-fold evaluation of the blend
- **Applies when**: `code` averages two or more models' test predictions, and/or uses a validation fold both for early stopping and for reporting that fold's score.
- **Pattern**: Per-fold XGBoost and LightGBM AUCs were printed,... | Ensemble weights and model choice fixed without any out-of-fold evaluation of the blend | Ensemble weights and model choice fixed without any out-of-fold evaluation of the blend | tabular-playground-series-may-2022 | tabular-playground-series-may-2022 | h2-mle-interp | 0ce5bc4eb40895cb | mle-bench-lite-LEAKED | mle-rubrics-oracle |
2f693ffabd23 | ### Only gradient-boosted trees tried, despite a task statement that the signal lives in feature interactions
- **Applies when**: `task` emphasizes feature interactions on a large tabular dataset (hundreds of thousands of rows) and `code` fits exclusively axis-aligned tree models.
- **Pattern**: The pipeline trains XGB... | Only gradient-boosted trees tried, despite a task statement that the signal lives in feature interactions | Only gradient-boosted trees tried, despite a task statement that the signal lives in feature interactions | tabular-playground-series-may-2022 | tabular-playground-series-may-2022 | h2-mle-interp | 0ce5bc4eb40895cb | mle-bench-lite-LEAKED | mle-rubrics-oracle |
c1c9bc2ffd6c | ### Hardcoded, unverified test-file path instead of discovering the actual data files
- **Applies when**: `code` -- the script reads competition inputs with literal paths such as `/home/data/ru_test_2.csv`, `/home/data/test.csv`, etc., and the competition description / instructions only guarantee generically named file... | Hardcoded, unverified test-file path instead of discovering the actual data files | Hardcoded, unverified test-file path instead of discovering the actual data files | text-normalization-challenge-russian-language | text-normalization-challenge-russian-language | h2-mle-interp | e4ad4280636df8ec | mle-bench-lite-LEAKED | mle-rubrics-oracle |
8c47f8dcb656 | ### All-or-nothing single script with no early fallback submission
- **Applies when**: `task` states a hard wall-clock/command budget and that "whatever file exists at /home/submission/submission.csv is graded AS IS", and `code` is one monolithic script that loads data, trains/builds a model, and writes the submission ... | All-or-nothing single script with no early fallback submission | All-or-nothing single script with no early fallback submission | text-normalization-challenge-russian-language | text-normalization-challenge-russian-language | h2-mle-interp | e4ad4280636df8ec | mle-bench-lite-LEAKED | mle-rubrics-oracle |
7f19afcc98bb | ### Submission never validated against the sample submission's ids, row count, and format
- **Applies when**: `code` constructs the submission by deriving ids itself (e.g. `sentence_id.astype(str) + '_' + token_id.astype(str)`) rather than by joining onto the provided sample submission.
- **Pattern**: The agent reconst... | Submission never validated against the sample submission's ids, row count, and format | Submission never validated against the sample submission's ids, row count, and format | text-normalization-challenge-russian-language | text-normalization-challenge-russian-language | h2-mle-interp | e4ad4280636df8ec | mle-bench-lite-LEAKED | mle-rubrics-oracle |
3a3d07950e1e | ### pandas default NA parsing silently corrupts literal text tokens
- **Applies when**: `task` is a token-level text task where the raw strings themselves may be `NA`, `nan`, `null`, `None`, `NaN`, `-`, or empty, and `code` loads them with a plain `pd.read_csv(...)`.
- **Pattern**: `read_csv` is called without `keep_de... | pandas default NA parsing silently corrupts literal text tokens | pandas default NA parsing silently corrupts literal text tokens | text-normalization-challenge-russian-language | text-normalization-challenge-russian-language | h2-mle-interp | e4ad4280636df8ec | mle-bench-lite-LEAKED | mle-rubrics-oracle |
cd4ae6d49314 | ### Pure memorization lookup with identity fallback and no rules for unseen semiotic classes
- **Applies when**: `task` requires normalizing open-ended classes (numbers, dates, times, currency, measures, transliteration) and `code` predicts solely via a `before -> most_common(after)` dictionary from train, falling back... | Pure memorization lookup with identity fallback and no rules for unseen semiotic classes | Pure memorization lookup with identity fallback and no rules for unseen semiotic classes | text-normalization-challenge-russian-language | text-normalization-challenge-russian-language | h2-mle-interp | e4ad4280636df8ec | mle-bench-lite-LEAKED | mle-rubrics-oracle |
End of preview. Expand in Data Studio
README.md exists but content is empty.
- Downloads last month
- 13