| { |
| "candidate_parameters": [ |
| "mass", |
| "drag coefficient", |
| "coefficient of restitution", |
| "gravity acceleration" |
| ], |
| "download_link": "https://huggingface.co/datasets/eth-siplab/tsenvbenchmark/tree/main/questions/BallDrop", |
| "environment_id": "BallDrop", |
| "name": "BallDrop", |
| "observed_channels": [ |
| { |
| "id": "Position", |
| "label": "ball height", |
| "unit": "" |
| }, |
| { |
| "id": "Velocity", |
| "label": "ball velocity", |
| "unit": "" |
| }, |
| { |
| "id": "Hard_Stop_f", |
| "label": "contact impulse", |
| "unit": "N*s" |
| } |
| ], |
| "prompt_combinations": [ |
| { |
| "agent_instruction": "The time series in the test_samples/ folder were generated by simulating a 1D vertical point-mass ball model under gravity.\nWhile the ball is above the ground, its motion is governed by gravity and quadratic air drag acting opposite the direction of travel, and the position evolves according to the current velocity.\nGround interaction is modeled with a restitution-based hard-stop law.\nWhen the ball reaches the ground with impact speed at or above 0.5 m/s, the impact is treated as instantaneous and the post-impact speed is set by the coefficient of restitution e in [0,1].\nThe rebound points upward and its magnitude equals e times the pre-impact speed.\nWhen the impact speed is below the threshold, the ball does not rebound and instead enters static contact at the ground.\n\nObserved Signals:\ncol1: ball height\ncol2: ball velocity\ncol3: contact impulse (N*s)\ncol4: time\n\nFor each simulation, either no parameter changes occur, or exactly one parameter among the allowed labels changes during the observed simulation interval.\nIf a parameter changes, it undergoes a single instantaneous step change at an unknown time during the observed interval.\n\nAllowed labels:\n[\"coefficient of restitution\", \"mass\", \"drag coefficient\", \"gravity acceleration\", \"no parameter change\"]\n\nUse \"no parameter change\" if there is no evidence in the data of a parameter change during the observed interval.\n\nTask:\nCreate a file named results.json in the current working directory.\nFor each file in test_samples/, return a ranked list of labels.\nThe first label is your final top-1 prediction and should be the single label you think is most likely correct.\nYou may include additional labels only when the evidence is genuinely ambiguous.\nAdditional labels are treated as lower-confidence alternatives.\nThe output must be valid JSON with exactly this structure:\n{\n\n \"<filename_1>\": [\"<top_1_label>\"],\n \"<filename_2>\": [\"<top_1_label>\", \"<optional_lower_confidence_label>\"]\n\n}\n\nRequirements:\n- Include one entry for every Parquet file in test_samples/.\n- Every returned label must exactly match one of the allowed labels.\n- The order of labels matters: the first label is the top-1 prediction.\n- Do not include duplicate labels for a sample.\n- Do not return all labels unless the evidence is genuinely ambiguous across all labels.\n\nEvaluation:\nThe primary evaluation metric is top-1 accuracy: the first label in each returned list is compared with the hidden correct label.\nA secondary shortlist score may also be reported. For a returned list of length m, the sample receives score 1/m if the hidden correct label appears anywhere in the list, and 0 otherwise. Therefore, unnecessary extra labels reduce the secondary score.\n\nAdditional requirements:\n- If you create intermediate files, images, scripts, or notes while solving the task, create them in the current working directory.\n- Internet access is disabled.", |
| "desc_level": "high", |
| "task_type": "direct", |
| "training_samples": "none" |
| }, |
| { |
| "agent_instruction": "The time series in the test_samples/ and train_samples/ folders were generated by simulating a 1D vertical point-mass ball model under gravity.\nWhile the ball is above the ground, its motion is governed by gravity and quadratic air drag acting opposite the direction of travel, and the position evolves according to the current velocity.\nGround interaction is modeled with a restitution-based hard-stop law.\nWhen the ball reaches the ground with impact speed at or above 0.5 m/s, the impact is treated as instantaneous and the post-impact speed is set by the coefficient of restitution e in [0,1].\nThe rebound points upward and its magnitude equals e times the pre-impact speed.\nWhen the impact speed is below the threshold, the ball does not rebound and instead enters static contact at the ground.\n\nObserved Signals:\ncol1: ball height\ncol2: ball velocity\ncol3: contact impulse (N*s)\ncol4: time\n\nFor each simulation, either no parameter changes occur, or exactly one parameter among the allowed labels changes during the observed simulation interval.\nIf a parameter changes, it undergoes a single instantaneous step change at an unknown time during the observed interval.\n\nAllowed labels:\n[\"coefficient of restitution\", \"mass\", \"drag coefficient\", \"gravity acceleration\", \"no parameter change\"]\n\nUse \"no parameter change\" if there is no evidence in the data of a parameter change during the observed interval.\n\nTask:\nCreate a file named results.json in the current working directory.\nFor each file in test_samples/, return a ranked list of labels.\nThe first label is your final top-1 prediction and should be the single label you think is most likely correct.\nYou may include additional labels only when the evidence is genuinely ambiguous.\nAdditional labels are treated as lower-confidence alternatives.\nThe output must be valid JSON with exactly this structure:\n{\n\n \"<filename_1>\": [\"<top_1_label>\"],\n \"<filename_2>\": [\"<top_1_label>\", \"<optional_lower_confidence_label>\"]\n\n}\n\nTo help with this task, you can use the labeled `train_samples/` directory. The corresponding labels are available in `train_labels.json` file.\n\nRequirements:\n- Include one entry for every Parquet file in test_samples/.\n- Every returned label must exactly match one of the allowed labels.\n- The order of labels matters: the first label is the top-1 prediction.\n- Do not include duplicate labels for a sample.\n- Do not return all labels unless the evidence is genuinely ambiguous across all labels.\n\nEvaluation:\nThe primary evaluation metric is top-1 accuracy: the first label in each returned list is compared with the hidden correct label.\nA secondary shortlist score may also be reported. For a returned list of length m, the sample receives score 1/m if the hidden correct label appears anywhere in the list, and 0 otherwise. Therefore, unnecessary extra labels reduce the secondary score.\n\nAdditional requirements:\n- If you create intermediate files, images, scripts, or notes while solving the task, create them in the current working directory.\n- Internet access is disabled.", |
| "desc_level": "high", |
| "task_type": "direct", |
| "training_samples": ">0" |
| }, |
| { |
| "agent_instruction": "Context:\nThe time series in the test_samples/ folder were generated by simulating a 1D vertical point-mass ball model under gravity.\nWhile the ball is above the ground, its motion is governed by gravity and quadratic air drag acting opposite the direction of travel, and the position evolves according to the current velocity.\nGround interaction is modeled with a restitution-based hard-stop law.\nWhen the ball reaches the ground with impact speed at or above 0.5 m/s, the impact is treated as instantaneous and the post-impact speed is set by the coefficient of restitution e in [0,1].\nThe rebound points upward and its magnitude equals e times the pre-impact speed.\nWhen the impact speed is below the threshold, the ball does not rebound and instead enters static contact at the ground.\n\nObserved Signals:\ncol1: ball height\ncol2: ball velocity\ncol3: contact impulse (N*s)\ncol4: time\n\nFor each simulation, either no parameter changes occur, or exactly one parameter among the allowed labels changes during the observed simulation interval.\nIf a parameter changes, it undergoes a single instantaneous step change at an unknown time during the observed interval.\n\nAllowed labels:\n[\"coefficient of restitution\", \"mass\", \"drag coefficient\", \"gravity acceleration\", \"no parameter change\"]\n\nUse \"no parameter change\" if there is no evidence in the data of a parameter change during the observed interval.\n\nTask:\nCreate a Python script named rule.py in the current working directory.\nThe script must define exactly this function:\ndef predict(df) -> list[str]:\nThe input df is a pandas DataFrame containing one sample with columns col1, col2, col3, and col4.\nThe function will be called on samples inside test_samples/ and on additional held-out samples with the same schema and label set.\nFor each dataframe, predict(df) must return a ranked list of labels. The first label is the final top-1 prediction and should be the single label most likely to be correct. Additional labels are optional lower-confidence alternatives and should only be included when the evidence is genuinely ambiguous.\n\nRequirements for predict(df):\n- Return a Python list of strings.\n- Every returned label must exactly match one of the allowed labels.\n- The first returned label is the top-1 prediction.\n- Do not include duplicate labels.\n- Do not return all labels unless the evidence is genuinely ambiguous across all labels.\n- You may inspect any file while developing rule.py, but the final submitted rule.py must not read, open, import, or depend on any files at prediction time, including test_samples/\n- The final rule.py must be able to run on a dataframe alone.\n\nEvaluation:\nThe primary evaluation metric is top-1 accuracy: the first label returned by predict(df) is compared with the hidden correct label.\nA secondary shortlist score may also be reported. For a returned list of length m, the sample receives score 1/m if the hidden correct label appears anywhere in the list, and 0 otherwise. Therefore, unnecessary extra labels reduce the secondary score.\n\nAdditional requirements:\n- If you create intermediate files, images, scripts, or notes while solving the task, create them in the current working directory.\n- Internet access is disabled.", |
| "desc_level": "high", |
| "task_type": "code", |
| "training_samples": "none" |
| }, |
| { |
| "agent_instruction": "Context:\nThe time series in the test_samples/ and train_samples/ folders were generated by simulating a 1D vertical point-mass ball model under gravity.\nWhile the ball is above the ground, its motion is governed by gravity and quadratic air drag acting opposite the direction of travel, and the position evolves according to the current velocity.\nGround interaction is modeled with a restitution-based hard-stop law.\nWhen the ball reaches the ground with impact speed at or above 0.5 m/s, the impact is treated as instantaneous and the post-impact speed is set by the coefficient of restitution e in [0,1].\nThe rebound points upward and its magnitude equals e times the pre-impact speed.\nWhen the impact speed is below the threshold, the ball does not rebound and instead enters static contact at the ground.\n\nObserved Signals:\ncol1: ball height\ncol2: ball velocity\ncol3: contact impulse (N*s)\ncol4: time\n\nFor each simulation, either no parameter changes occur, or exactly one parameter among the allowed labels changes during the observed simulation interval.\nIf a parameter changes, it undergoes a single instantaneous step change at an unknown time during the observed interval.\n\nAllowed labels:\n[\"coefficient of restitution\", \"mass\", \"drag coefficient\", \"gravity acceleration\", \"no parameter change\"]\n\nUse \"no parameter change\" if there is no evidence in the data of a parameter change during the observed interval.\n\nTask:\nCreate a Python script named rule.py in the current working directory.\nThe script must define exactly this function:\ndef predict(df) -> list[str]:\nThe input df is a pandas DataFrame containing one sample with columns col1, col2, col3, and col4.\nThe function will be called on samples inside test_samples/ and on additional held-out samples with the same schema and label set.\nFor each dataframe, predict(df) must return a ranked list of labels. The first label is the final top-1 prediction and should be the single label most likely to be correct. Additional labels are optional lower-confidence alternatives and should only be included when the evidence is genuinely ambiguous.\n\nTo help with this task, you can use the labeled train_samples/ directory while developing rule.py. The corresponding labels are available in train_labels.json.\n\nRequirements for predict(df):\n- Return a Python list of strings.\n- Every returned label must exactly match one of the allowed labels.\n- The first returned label is the top-1 prediction.\n- Do not include duplicate labels.\n- Do not return all labels unless the evidence is genuinely ambiguous across all labels.\n- You may inspect train_samples/ and train_labels.json while developing rule.py, but the final submitted rule.py must not read, open, import, or depend on any files at prediction time, including train_samples/, test_samples/ or train_labels.json\n- The final rule.py must be able to run on a dataframe alone.\n\nEvaluation:\nThe primary evaluation metric is top-1 accuracy: the first label returned by predict(df) is compared with the hidden correct label.\nA secondary shortlist score may also be reported. For a returned list of length m, the sample receives score 1/m if the hidden correct label appears anywhere in the list, and 0 otherwise. Therefore, unnecessary extra labels reduce the secondary score.\n\nAdditional requirements:\n- If you create intermediate files, images, scripts, or notes while solving the task, create them in the current working directory.\n- Internet access is disabled.", |
| "desc_level": "high", |
| "task_type": "code", |
| "training_samples": ">0" |
| }, |
| { |
| "agent_instruction": "Context:\nThe time series in the test_samples/ folder were generated by a simulator of an unknown physical phenomenon.\nThe column meanings are unknown, except for the last column, which represents time.\n\nFor each simulation, either no parameter changes occur, or exactly one parameter corresponding to one of the allowed changes during the observed simulation interval.\nIf a parameter changes, it undergoes a single instantaneous step change at an unknown time during the observed interval.\n\n[\"label_0\", \"label_1\", \"label_2\", \"label_3\"] denote different parameter changes, while \"label_4\" denotes that no parameter changed.\n\nTask:\nCreate a file named results.json in the current working directory.\nFor each file in test_samples/, return a ranked list of labels.\nThe first label is your final top-1 prediction and should be the single label you think is most likely correct.\nYou may include additional labels only when the evidence is genuinely ambiguous.\nAdditional labels are treated as lower-confidence alternatives.\nThe output must be valid JSON with exactly this structure:\n{\n\n \"<filename_1>\": [\"<top_1_label>\"],\n \"<filename_2>\": [\"<top_1_label>\", \"<optional_lower_confidence_label>\"]\n\n}\n\nRequirements:\n- Include one entry for every Parquet file in test_samples/.\n- Every returned label must exactly match one of the allowed labels.\n- The order of labels matters: the first label is the top-1 prediction.\n- Do not include duplicate labels for a sample.\n- Do not return all labels unless the evidence is genuinely ambiguous across all labels.\n\nEvaluation:\nThe primary evaluation metric is top-1 accuracy: the first label in each returned list is compared with the hidden correct label.\nA secondary shortlist score may also be reported. For a returned list of length m, the sample receives score 1/m if the hidden correct label appears anywhere in the list, and 0 otherwise. Therefore, unnecessary extra labels reduce the secondary score.\n\nAdditional requirements:\n- If you create intermediate files, images, scripts, or notes while solving the task, create them in the current working directory.\n- Internet access is disabled.", |
| "desc_level": "none", |
| "task_type": "direct", |
| "training_samples": "none" |
| }, |
| { |
| "agent_instruction": "Context:\nThe time series in the test_samples/ and train_samples/ folders were generated by a simulator of an unknown physical phenomenon.\nThe column meanings are unknown, except for the last column, which represents time.\n\nFor each simulation, either no parameter changes occur, or exactly one parameter corresponding to one of the allowed changes during the observed simulation interval.\nIf a parameter changes, it undergoes a single instantaneous step change at an unknown time during the observed interval.\n\n[\"label_0\", \"label_1\", \"label_2\", \"label_3\"] denote different parameter changes, while \"label_4\" denotes that no parameter changed.\n\nTask:\nCreate a file named results.json in the current working directory.\nFor each file in test_samples/, return a ranked list of labels.\nThe first label is your final top-1 prediction and should be the single label you think is most likely correct.\nYou may include additional labels only when the evidence is genuinely ambiguous.\nAdditional labels are treated as lower-confidence alternatives.\nThe output must be valid JSON with exactly this structure:\n{\n\n \"<filename_1>\": [\"<top_1_label>\"],\n \"<filename_2>\": [\"<top_1_label>\", \"<optional_lower_confidence_label>\"]\n\n}\n\nTo help with this task, you can use the labeled `train_samples/` directory. The corresponding labels are available in `train_labels.json` file.\n\nRequirements:\n- Include one entry for every Parquet file in test_samples/.\n- Every returned label must exactly match one of the allowed labels.\n- The order of labels matters: the first label is the top-1 prediction.\n- Do not include duplicate labels for a sample.\n- Do not return all labels unless the evidence is genuinely ambiguous across all labels.\n\nEvaluation:\nThe primary evaluation metric is top-1 accuracy: the first label in each returned list is compared with the hidden correct label.\nA secondary shortlist score may also be reported. For a returned list of length m, the sample receives score 1/m if the hidden correct label appears anywhere in the list, and 0 otherwise. Therefore, unnecessary extra labels reduce the secondary score.\n\nAdditional requirements:\n- If you create intermediate files, images, scripts, or notes while solving the task, create them in the current working directory.\n- Internet access is disabled.", |
| "desc_level": "none", |
| "task_type": "direct", |
| "training_samples": ">0" |
| }, |
| { |
| "agent_instruction": "Context:\nThe time series in the test_samples/ folder were generated by a simulator of an unknown physical phenomenon.\nThe column meanings are unknown, except for the last column, which represents time.\n\nFor each simulation, either no parameter changes occur, or exactly one parameter corresponding to one of the allowed changes during the observed simulation interval.\nIf a parameter changes, it undergoes a single instantaneous step change at an unknown time during the observed interval.\n\n[\"label_0\", \"label_1\", \"label_2\", \"label_3\"] denote different parameter changes, while \"label_4\" denotes that no parameter changed.\n\nTask:\nCreate a Python script named rule.py in the current working directory.\nThe script must define exactly this function:\ndef predict(df) -> list[str]:\nThe input df is a pandas DataFrame containing one sample with columns col1, col2, col3, and col4.\nThe function will be called on samples inside test_samples/ and on additional held-out samples with the same schema and label set.\nFor each dataframe, predict(df) must return a ranked list of labels. The first label is the final top-1 prediction and should be the single label most likely to be correct. Additional labels are optional lower-confidence alternatives and should only be included when the evidence is genuinely ambiguous.\n\nRequirements for predict(df):\n- Return a Python list of strings.\n- Every returned label must exactly match one of the allowed labels.\n- The first returned label is the top-1 prediction.\n- Do not include duplicate labels.\n- Do not return all labels unless the evidence is genuinely ambiguous across all labels.\n- You may inspect any file while developing rule.py, but the final submitted rule.py must not read, open, import, or depend on any files at prediction time, including test_samples/\n- The final rule.py must be able to run on a dataframe alone.\n\nEvaluation:\nThe primary evaluation metric is top-1 accuracy: the first label returned by predict(df) is compared with the hidden correct label.\nA secondary shortlist score may also be reported. For a returned list of length m, the sample receives score 1/m if the hidden correct label appears anywhere in the list, and 0 otherwise. Therefore, unnecessary extra labels reduce the secondary score.\n\nAdditional requirements:\n- If you create intermediate files, images, scripts, or notes while solving the task, create them in the current working directory.\n- Internet access is disabled.", |
| "desc_level": "none", |
| "task_type": "code", |
| "training_samples": "none" |
| }, |
| { |
| "agent_instruction": "Context:\nThe time series in the test_samples/ and train_samples/ folders were generated by a simulator of an unknown physical phenomenon.\nThe column meanings are unknown, except for the last column, which represents time.\n\nFor each simulation, either no parameter changes occur, or exactly one parameter corresponding to one of the allowed changes during the observed simulation interval.\nIf a parameter changes, it undergoes a single instantaneous step change at an unknown time during the observed interval.\n\n[\"label_0\", \"label_1\", \"label_2\", \"label_3\"] denote different parameter changes, while \"label_4\" denotes that no parameter changed.\n\nTask:\nCreate a Python script named rule.py in the current working directory.\nThe script must define exactly this function:\ndef predict(df) -> list[str]:\nThe input df is a pandas DataFrame containing one sample with columns col1, col2, col3, and col4.\nThe function will be called on samples inside test_samples/ and on additional held-out samples with the same schema and label set.\nFor each dataframe, predict(df) must return a ranked list of labels. The first label is the final top-1 prediction and should be the single label most likely to be correct. Additional labels are optional lower-confidence alternatives and should only be included when the evidence is genuinely ambiguous.\n\nTo help with this task, you can use the labeled train_samples/ directory while developing rule.py. The corresponding labels are available in train_labels.json.\n\nRequirements for predict(df):\n- Return a Python list of strings.\n- Every returned label must exactly match one of the allowed labels.\n- The first returned label is the top-1 prediction.\n- Do not include duplicate labels.\n- Do not return all labels unless the evidence is genuinely ambiguous across all labels.\n- You may inspect train_samples/ and train_labels.json while developing rule.py, but the final submitted rule.py must not read, open, import, or depend on any files at prediction time, including train_samples/, test_samples/ or train_labels.json\n- The final rule.py must be able to run on a dataframe alone.\n\nEvaluation:\nThe primary evaluation metric is top-1 accuracy: the first label returned by predict(df) is compared with the hidden correct label.\nA secondary shortlist score may also be reported. For a returned list of length m, the sample receives score 1/m if the hidden correct label appears anywhere in the list, and 0 otherwise. Therefore, unnecessary extra labels reduce the secondary score.\n\nAdditional requirements:\n- If you create intermediate files, images, scripts, or notes while solving the task, create them in the current working directory.\n- Internet access is disabled.", |
| "desc_level": "none", |
| "task_type": "code", |
| "training_samples": ">0" |
| } |
| ], |
| "sample_count": 5, |
| "short_one_line_description": "A ball is dropped from a height and observed over time." |
| } |
|
|
