{
  "id": 61455,
  "title": "Data issue (funny tracks) explained",
  "url": "/competitions/trackml-particle-identification/discussion/61455",
  "author_name": "",
  "post_date": "2018-07-19T14:36:39.010860200Z",
  "votes": 6,
  "comment_count": 7,
  "views": 0,
  "content": "<p>Two issues have been identified in the dataset, both affecting only a small fraction of particles and within those mostly particles with a very low score (due to their kinematic properties).\nDetails about the impact on the score can be found below.</p>\n\n<p>For Phase 2 of the challenge, we will fix these issues.</p>\n\n<h3>We apologize for these issues and hope they did not create too much inconvenience.</h3>\n\n<h1>Non-physical tracks</h1>\n\n<h2>Description</h2>\n\n<p>A non-physical scattering description has been found for electron particles in the dataset. As electrons have a significantly lower mass than all other charged particles, a different scattering formula is applied. In practice, the so-called Highland formula is often replaced by another formalism (Rossi-Greisen) that modifies the Highland formula slightly.</p>\n\n<p>In the simulation code, this leads to different constants for electrons and all other particles. In the simulation module for electrons used for the dataset production a faulty unit (MeV instead of GeV) was used. This overestimates the scattering by a factor of thousand, leading to non-physical particle trajectories: while the particle moves through the magnetic field as a GeV scale particle, the material interaction is applied as if it were a MeV scale particle. A low-energy MeV scale particle scatters significantly more than a high-energy particles, thus leading to the broken trajectories found in the dataset.</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10065/Fraction.png\" alt=\"Fraction of electrons in the dataset for several test events.\"></p>\n\n<p>Fortunately, as shown in the figure above, the fraction of electrons in the dataset is small and the fraction of hits generated by them is even smaller. Electrons tend to occur as secondary particles with low momentum and a lower average number of hits. The transverse momentum distribution is also shifted to lower values leading to a reduced impact on the score as a result of the transverse momentum dependant hit score, as shown in the figure below. First tests evaluating the impact on the score for existing submissions show negligible impact compared to the overall score.</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10068/Weight.png\" alt=\"Kinematic properties of electrons compared to all charged particles (left) and track weights by momentum (right).\"></p>\n\n<h2>Impact evaluation on score</h2>\n\n<p>The following graph shows the original score and a corrected score for several test submissions, when the affected particles are removed from the scoring.</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10066/ScoreChange.png\" alt=\"Change of score when electrons are removed.\"></p>\n\n<p>As expected, there is a slight shift of the score upwards as these particles are practically not to find due to their random behavior. \nThus, by removing them from the score pool, the submitted solutions gain a fractional increase of about 0.16 % which is consistent with the fraction of hits created by electrons to the total hit content used in the scoring.</p>\n\n<h3>A blacklist of hits and particle ids that originate from these electrons has been created and will be uploaded shortly to the competition page.</h3>\n\n<p>This can be done both for the test and for the training dataset. They only contribute a minor fraction to the score compared to the current best submissions and can be safely ignored without impacting the leaderboard results. Everything else, including the score currently implemented by Kaggle stays the same.</p>\n\n<h1>Loopers</h1>\n\n<h2>Description</h2>\n\n<p>Looping particles are particles that have such a low momentum that bend fully within the detector volume and thus can exit the detector only along the longitudinal direction (they \"spiral\" out of the detector if not being destroyed by interaction with the detector material). If the particle has almost no longitudinal momentum direction, however, they can stay within the central region of the detector for a rather long time.</p>\n\n<p>They exist in reality and add to the confusion of pattern recognition. For certain cases,  this is indeed happening in the dataset, with the requirement that the particles do not exist the virtual tracking volume (we stop particles at the detector boundary), you can see some examples below.</p>\n\n<p>The hits created by loopers later along their path are indeed \"on track\", i.e. on the physical movement on track, albeit they may seem to be detached from the trajectory. The used fast simulation engine is not designed to handle those loopers with ultimate precision, resulting in potential lower hit efficiencies for those tracks.</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10069/Loopers.png\" alt=\"Display of some tracks from loopers with actually progressed path and hit description.\">\n![Display of some tracks from loopers with actually progressed path and hit description.]</p>\n\n<h2>Impact evaluation</h2>\n\n<p>Similarly to the <em>non-physical</em> tracks we will evaluate the score impact of those loopers, the creation of a potential black list needs re-simulation with a  dedicated loop protection, we will investigate.</p>",
  "messages": [
    {
      "id": "359151",
      "postDate": "07/19/2018 14:36:39",
      "content": "<p>Two issues have been identified in the dataset, both affecting only a small fraction of particles and within those mostly particles with a very low score (due to their kinematic properties).\nDetails about the impact on the score can be found below.</p>\n\n<p>For Phase 2 of the challenge, we will fix these issues.</p>\n\n<h3>We apologize for these issues and hope they did not create too much inconvenience.</h3>\n\n<h1>Non-physical tracks</h1>\n\n<h2>Description</h2>\n\n<p>A non-physical scattering description has been found for electron particles in the dataset. As electrons have a significantly lower mass than all other charged particles, a different scattering formula is applied. In practice, the so-called Highland formula is often replaced by another formalism (Rossi-Greisen) that modifies the Highland formula slightly.</p>\n\n<p>In the simulation code, this leads to different constants for electrons and all other particles. In the simulation module for electrons used for the dataset production a faulty unit (MeV instead of GeV) was used. This overestimates the scattering by a factor of thousand, leading to non-physical particle trajectories: while the particle moves through the magnetic field as a GeV scale particle, the material interaction is applied as if it were a MeV scale particle. A low-energy MeV scale particle scatters significantly more than a high-energy particles, thus leading to the broken trajectories found in the dataset.</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10065/Fraction.png\" alt=\"Fraction of electrons in the dataset for several test events.\"></p>\n\n<p>Fortunately, as shown in the figure above, the fraction of electrons in the dataset is small and the fraction of hits generated by them is even smaller. Electrons tend to occur as secondary particles with low momentum and a lower average number of hits. The transverse momentum distribution is also shifted to lower values leading to a reduced impact on the score as a result of the transverse momentum dependant hit score, as shown in the figure below. First tests evaluating the impact on the score for existing submissions show negligible impact compared to the overall score.</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10068/Weight.png\" alt=\"Kinematic properties of electrons compared to all charged particles (left) and track weights by momentum (right).\"></p>\n\n<h2>Impact evaluation on score</h2>\n\n<p>The following graph shows the original score and a corrected score for several test submissions, when the affected particles are removed from the scoring.</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10066/ScoreChange.png\" alt=\"Change of score when electrons are removed.\"></p>\n\n<p>As expected, there is a slight shift of the score upwards as these particles are practically not to find due to their random behavior. \nThus, by removing them from the score pool, the submitted solutions gain a fractional increase of about 0.16 % which is consistent with the fraction of hits created by electrons to the total hit content used in the scoring.</p>\n\n<h3>A blacklist of hits and particle ids that originate from these electrons has been created and will be uploaded shortly to the competition page.</h3>\n\n<p>This can be done both for the test and for the training dataset. They only contribute a minor fraction to the score compared to the current best submissions and can be safely ignored without impacting the leaderboard results. Everything else, including the score currently implemented by Kaggle stays the same.</p>\n\n<h1>Loopers</h1>\n\n<h2>Description</h2>\n\n<p>Looping particles are particles that have such a low momentum that bend fully within the detector volume and thus can exit the detector only along the longitudinal direction (they \"spiral\" out of the detector if not being destroyed by interaction with the detector material). If the particle has almost no longitudinal momentum direction, however, they can stay within the central region of the detector for a rather long time.</p>\n\n<p>They exist in reality and add to the confusion of pattern recognition. For certain cases,  this is indeed happening in the dataset, with the requirement that the particles do not exist the virtual tracking volume (we stop particles at the detector boundary), you can see some examples below.</p>\n\n<p>The hits created by loopers later along their path are indeed \"on track\", i.e. on the physical movement on track, albeit they may seem to be detached from the trajectory. The used fast simulation engine is not designed to handle those loopers with ultimate precision, resulting in potential lower hit efficiencies for those tracks.</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10069/Loopers.png\" alt=\"Display of some tracks from loopers with actually progressed path and hit description.\">\n![Display of some tracks from loopers with actually progressed path and hit description.]</p>\n\n<h2>Impact evaluation</h2>\n\n<p>Similarly to the <em>non-physical</em> tracks we will evaluate the score impact of those loopers, the creation of a potential black list needs re-simulation with a  dedicated loop protection, we will investigate.</p>",
      "rawMarkdown": "Two issues have been identified in the dataset, both affecting only a small fraction of particles and within those mostly particles with a very low score (due to their kinematic properties).\nDetails about the impact on the score can be found below.\n\nFor Phase 2 of the challenge, we will fix these issues.\n\n###We apologize for these issues and hope they did not create too much inconvenience.\n\n#Non-physical tracks\n\n##Description\n\nA non-physical scattering description has been found for electron particles in the dataset. As electrons have a significantly lower mass than all other charged particles, a different scattering formula is applied. In practice, the so-called Highland formula is often replaced by another formalism (Rossi-Greisen) that modifies the Highland formula slightly.\n\nIn the simulation code, this leads to different constants for electrons and all other particles. In the simulation module for electrons used for the dataset production a faulty unit (MeV instead of GeV) was used. This overestimates the scattering by a factor of thousand, leading to non-physical particle trajectories: while the particle moves through the magnetic field as a GeV scale particle, the material interaction is applied as if it were a MeV scale particle. A low-energy MeV scale particle scatters significantly more than a high-energy particles, thus leading to the broken trajectories found in the dataset.\n\n![Fraction of electrons in the dataset for several test events.][1]\n\nFortunately, as shown in the figure above, the fraction of electrons in the dataset is small and the fraction of hits generated by them is even smaller. Electrons tend to occur as secondary particles with low momentum and a lower average number of hits. The transverse momentum distribution is also shifted to lower values leading to a reduced impact on the score as a result of the transverse momentum dependant hit score, as shown in the figure below. First tests evaluating the impact on the score for existing submissions show negligible impact compared to the overall score.\n\n![Kinematic properties of electrons compared to all charged particles (left) and track weights by momentum (right).][2]\n\n##Impact evaluation on score\n\nThe following graph shows the original score and a corrected score for several test submissions, when the affected particles are removed from the scoring.\n\n![Change of score when electrons are removed.][3]\n\nAs expected, there is a slight shift of the score upwards as these particles are practically not to find due to their random behavior. \nThus, by removing them from the score pool, the submitted solutions gain a fractional increase of about 0.16 % which is consistent with the fraction of hits created by electrons to the total hit content used in the scoring.\n\n###A blacklist of hits and particle ids that originate from these electrons has been created and will be uploaded shortly to the competition page.\n\nThis can be done both for the test and for the training dataset. They only contribute a minor fraction to the score compared to the current best submissions and can be safely ignored without impacting the leaderboard results. Everything else, including the score currently implemented by Kaggle stays the same.\n\n#Loopers\n\n##Description\n\nLooping particles are particles that have such a low momentum that bend fully within the detector volume and thus can exit the detector only along the longitudinal direction (they \"spiral\" out of the detector if not being destroyed by interaction with the detector material). If the particle has almost no longitudinal momentum direction, however, they can stay within the central region of the detector for a rather long time.\n\nThey exist in reality and add to the confusion of pattern recognition. For certain cases,  this is indeed happening in the dataset, with the requirement that the particles do not exist the virtual tracking volume (we stop particles at the detector boundary), you can see some examples below.\n\nThe hits created by loopers later along their path are indeed \"on track\", i.e. on the physical movement on track, albeit they may seem to be detached from the trajectory. The used fast simulation engine is not designed to handle those loopers with ultimate precision, resulting in potential lower hit efficiencies for those tracks.\n\n![Display of some tracks from loopers with actually progressed path and hit description.][4]\n![Display of some tracks from loopers with actually progressed path and hit description.]\n\n##Impact evaluation \n\nSimilarly to the *non-physical* tracks we will evaluate the score impact of those loopers, the creation of a potential black list needs re-simulation with a  dedicated loop protection, we will investigate.\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10065/Fraction.png\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10068/Weight.png\n  [3]: https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10066/ScoreChange.png\n  [4]: https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10069/Loopers.png",
      "votes": null
    },
    {
      "id": "359210",
      "postDate": "07/19/2018 16:49:49",
      "content": "<p>HI, thanks for coming back. It seems you identified several classes of issues, but, if I read correctly all your posts from today, the 'bouncing particles' are still not explained.  Do I get it right?  </p>\n\n<p>Re the electron scattering, will you provide the list of all impacted hits for all train and test data? Sure, the impact on score is not that large, but these particles have low weights hit, hence the number of affected hits may be significant larger than 0.16%.  </p>",
      "rawMarkdown": "HI, thanks for coming back. It seems you identified several classes of issues, but, if I read correctly all your posts from today, the 'bouncing particles' are still not explained.  Do I get it right?  \n\nRe the electron scattering, will you provide the list of all impacted hits for all train and test data? Sure, the impact on score is not that large, but these particles have low weights hit, hence the number of affected hits may be significant larger than 0.16%.",
      "votes": null
    },
    {
      "id": "359282",
      "postDate": "07/19/2018 19:51:35",
      "content": "<p>Hi, the total fraction of hits impacted by the electron scattering is about 0.14%. We will provide a list of impacted particles and hits that should become available soon.</p>",
      "rawMarkdown": "Hi, the total fraction of hits impacted by the electron scattering is about 0.14%. We will provide a list of impacted particles and hits that should become available soon.",
      "votes": null
    },
    {
      "id": "359506",
      "postDate": "07/20/2018 08:28:08",
      "content": "<p>Rereading this, the particle in the last picture seems to bounce. There is a first series of circles on the left, along one helix, then a bounce, and the circles go on the right onto another helix.  Is electron scattering causing this significant change of direction, or is it a yet to be found bug?</p>",
      "rawMarkdown": "Rereading this, the particle in the last picture seems to bounce. There is a first series of circles on the left, along one helix, then a bounce, and the circles go on the right onto another helix.  Is electron scattering causing this significant change of direction, or is it a yet to be found bug?",
      "votes": null
    },
    {
      "id": "359597",
      "postDate": "07/20/2018 12:22:54",
      "content": "<p>The picture is actually not of one particle but of several starting from the same position with +/- random charge (this was for me to find a reproducible candidate) - this is why you see the V structure here.</p>\n\n<p>For clarification, I did split out the track (I can't do that with the trajectories at the moment that easily), so you will see the trajectory of all my tests with one track:</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10067/Looper.png\" alt=\"One track embedded into the trajectory plots of 5 test particles.\"></p>",
      "rawMarkdown": "The picture is actually not of one particle but of several starting from the same position with +/- random charge (this was for me to find a reproducible candidate) - this is why you see the V structure here.\n\nFor clarification, I did split out the track (I can't do that with the trajectories at the moment that easily), so you will see the trajectory of all my tests with one track:\n\n![One track embedded into the trajectory plots of 5 test particles.][1]\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10067/Looper.png",
      "votes": null
    },
    {
      "id": "359731",
      "postDate": "07/20/2018 17:11:41",
      "content": "<p>Thanks.</p>",
      "rawMarkdown": "Thanks.",
      "votes": null
    },
    {
      "id": "362163",
      "postDate": "07/25/2018 20:01:29",
      "content": "<p>Hi everyone,</p>\n\n<p>the blacklist files are now available on the data page as <code>blacklist_training.zip</code>. For each event in the training dataset, it contains the list of particle ids for electrons that exhibit the wrong scattering behaviour as well as the hit ids generated by these specific particles.</p>",
      "rawMarkdown": "Hi everyone,\n\nthe blacklist files are now available on the data page as `blacklist_training.zip`. For each event in the training dataset, it contains the list of particle ids for electrons that exhibit the wrong scattering behaviour as well as the hit ids generated by these specific particles.",
      "votes": null
    },
    {
      "id": "362182",
      "postDate": "07/25/2018 21:12:44",
      "content": "<p>Thank you!\nAre you going to release the blacklist also for the test set?</p>",
      "rawMarkdown": "Thank you!\nAre you going to release the blacklist also for the test set?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 359210,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "07/19/2018 16:49:49",
      "content": "<p>HI, thanks for coming back. It seems you identified several classes of issues, but, if I read correctly all your posts from today, the 'bouncing particles' are still not explained.  Do I get it right?  </p>\n\n<p>Re the electron scattering, will you provide the list of all impacted hits for all train and test data? Sure, the impact on score is not that large, but these particles have low weights hit, hence the number of affected hits may be significant larger than 0.16%.  </p>",
      "votes": null,
      "replies": [
        {
          "id": 359282,
          "author_name": "msmk00",
          "author_url": "",
          "post_date": "07/19/2018 19:51:35",
          "content": "<p>Hi, the total fraction of hits impacted by the electron scattering is about 0.14%. We will provide a list of impacted particles and hits that should become available soon.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 359506,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "07/20/2018 08:28:08",
      "content": "<p>Rereading this, the particle in the last picture seems to bounce. There is a first series of circles on the left, along one helix, then a bounce, and the circles go on the right onto another helix.  Is electron scattering causing this significant change of direction, or is it a yet to be found bug?</p>",
      "votes": null,
      "replies": [
        {
          "id": 359597,
          "author_name": "asalzburger",
          "author_url": "",
          "post_date": "07/20/2018 12:22:54",
          "content": "<p>The picture is actually not of one particle but of several starting from the same position with +/- random charge (this was for me to find a reproducible candidate) - this is why you see the V structure here.</p>\n\n<p>For clarification, I did split out the track (I can't do that with the trajectories at the moment that easily), so you will see the trajectory of all my tests with one track:</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10067/Looper.png\" alt=\"One track embedded into the trajectory plots of 5 test particles.\"></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 359731,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "07/20/2018 17:11:41",
          "content": "<p>Thanks.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 362163,
      "author_name": "msmk00",
      "author_url": "",
      "post_date": "07/25/2018 20:01:29",
      "content": "<p>Hi everyone,</p>\n\n<p>the blacklist files are now available on the data page as <code>blacklist_training.zip</code>. For each event in the training dataset, it contains the list of particle ids for electrons that exhibit the wrong scattering behaviour as well as the hit ids generated by these specific particles.</p>",
      "votes": null,
      "replies": [
        {
          "id": 362182,
          "author_name": "panosc",
          "author_url": "",
          "post_date": "07/25/2018 21:12:44",
          "content": "<p>Thank you!\nAre you going to release the blacklist also for the test set?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "359151": "Two issues have been identified in the dataset, both affecting only a small fraction of particles and within those mostly particles with a very low score (due to their kinematic properties).\nDetails about the impact on the score can be found below.\n\nFor Phase 2 of the challenge, we will fix these issues.\n\n###We apologize for these issues and hope they did not create too much inconvenience.\n\n#Non-physical tracks\n\n##Description\n\nA non-physical scattering description has been found for electron particles in the dataset. As electrons have a significantly lower mass than all other charged particles, a different scattering formula is applied. In practice, the so-called Highland formula is often replaced by another formalism (Rossi-Greisen) that modifies the Highland formula slightly.\n\nIn the simulation code, this leads to different constants for electrons and all other particles. In the simulation module for electrons used for the dataset production a faulty unit (MeV instead of GeV) was used. This overestimates the scattering by a factor of thousand, leading to non-physical particle trajectories: while the particle moves through the magnetic field as a GeV scale particle, the material interaction is applied as if it were a MeV scale particle. A low-energy MeV scale particle scatters significantly more than a high-energy particles, thus leading to the broken trajectories found in the dataset.\n\n![Fraction of electrons in the dataset for several test events.][1]\n\nFortunately, as shown in the figure above, the fraction of electrons in the dataset is small and the fraction of hits generated by them is even smaller. Electrons tend to occur as secondary particles with low momentum and a lower average number of hits. The transverse momentum distribution is also shifted to lower values leading to a reduced impact on the score as a result of the transverse momentum dependant hit score, as shown in the figure below. First tests evaluating the impact on the score for existing submissions show negligible impact compared to the overall score.\n\n![Kinematic properties of electrons compared to all charged particles (left) and track weights by momentum (right).][2]\n\n##Impact evaluation on score\n\nThe following graph shows the original score and a corrected score for several test submissions, when the affected particles are removed from the scoring.\n\n![Change of score when electrons are removed.][3]\n\nAs expected, there is a slight shift of the score upwards as these particles are practically not to find due to their random behavior. \nThus, by removing them from the score pool, the submitted solutions gain a fractional increase of about 0.16 % which is consistent with the fraction of hits created by electrons to the total hit content used in the scoring.\n\n###A blacklist of hits and particle ids that originate from these electrons has been created and will be uploaded shortly to the competition page.\n\nThis can be done both for the test and for the training dataset. They only contribute a minor fraction to the score compared to the current best submissions and can be safely ignored without impacting the leaderboard results. Everything else, including the score currently implemented by Kaggle stays the same.\n\n#Loopers\n\n##Description\n\nLooping particles are particles that have such a low momentum that bend fully within the detector volume and thus can exit the detector only along the longitudinal direction (they \"spiral\" out of the detector if not being destroyed by interaction with the detector material). If the particle has almost no longitudinal momentum direction, however, they can stay within the central region of the detector for a rather long time.\n\nThey exist in reality and add to the confusion of pattern recognition. For certain cases,  this is indeed happening in the dataset, with the requirement that the particles do not exist the virtual tracking volume (we stop particles at the detector boundary), you can see some examples below.\n\nThe hits created by loopers later along their path are indeed \"on track\", i.e. on the physical movement on track, albeit they may seem to be detached from the trajectory. The used fast simulation engine is not designed to handle those loopers with ultimate precision, resulting in potential lower hit efficiencies for those tracks.\n\n![Display of some tracks from loopers with actually progressed path and hit description.][4]\n![Display of some tracks from loopers with actually progressed path and hit description.]\n\n##Impact evaluation \n\nSimilarly to the *non-physical* tracks we will evaluate the score impact of those loopers, the creation of a potential black list needs re-simulation with a  dedicated loop protection, we will investigate.\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10065/Fraction.png\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10068/Weight.png\n  [3]: https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10066/ScoreChange.png\n  [4]: https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10069/Loopers.png",
    "359210": "HI, thanks for coming back. It seems you identified several classes of issues, but, if I read correctly all your posts from today, the 'bouncing particles' are still not explained.  Do I get it right?  \n\nRe the electron scattering, will you provide the list of all impacted hits for all train and test data? Sure, the impact on score is not that large, but these particles have low weights hit, hence the number of affected hits may be significant larger than 0.16%.",
    "359282": "Hi, the total fraction of hits impacted by the electron scattering is about 0.14%. We will provide a list of impacted particles and hits that should become available soon.",
    "359506": "Rereading this, the particle in the last picture seems to bounce. There is a first series of circles on the left, along one helix, then a bounce, and the circles go on the right onto another helix.  Is electron scattering causing this significant change of direction, or is it a yet to be found bug?",
    "359597": "The picture is actually not of one particle but of several starting from the same position with +/- random charge (this was for me to find a reproducible candidate) - this is why you see the V structure here.\n\nFor clarification, I did split out the track (I can't do that with the trajectories at the moment that easily), so you will see the trajectory of all my tests with one track:\n\n![One track embedded into the trajectory plots of 5 test particles.][1]\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/359151/10067/Looper.png",
    "359731": "Thanks.",
    "362163": "Hi everyone,\n\nthe blacklist files are now available on the data page as `blacklist_training.zip`. For each event in the training dataset, it contains the list of particle ids for electrons that exhibit the wrong scattering behaviour as well as the hit ids generated by these specific particles.",
    "362182": "Thank you!\nAre you going to release the blacklist also for the test set?"
  },
  "source": "meta"
}