{
  "id": 680261,
  "title": "Score Breakdown of All Finalist Teams ",
  "url": "/competitions/dl-sprint-4-0-bengali-long-form-speech-recognition/discussion/680261",
  "author_name": "Shadman Tabib",
  "post_date": "2026-03-06T20:31:04.334000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<h1>DL Sprint 4.0 — Final Scores &amp; Evaluation Breakdown</h1>\n<p>Congratulations to all the teams who made it to the top 18 — and an even bigger congratulations to every single participant who gave it their best. Reaching the final round out of <strong>192 registered teams</strong>, while competing in two intense competitions simultaneously, is no small feat. Regardless of where you placed, the effort, learning, and perseverance you demonstrated throughout this sprint speak for themselves.</p>\n<p>Best of luck to all of you in your future endeavours. </p>\n<p>In the spirit of transparency, we are making the complete score breakdown public so that every team can see exactly how the final rankings were determined.</p>\n<hr>\n<table>\n<thead>\n<tr>\n<th>Rank</th>\n<th>Team</th>\n<th>Net Online Score</th>\n<th>Net Offline Score</th>\n<th>Final Weighted Score</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>1</td>\n<td>MaraKheyeGechi</td>\n<td>0.8687789541</td>\n<td>0.80500</td>\n<td>0.8496452679</td>\n</tr>\n<tr>\n<td>2</td>\n<td>Labyrinth</td>\n<td>0.8339024755</td>\n<td>0.87625</td>\n<td>0.8466067328</td>\n</tr>\n<tr>\n<td>3</td>\n<td>Villagers</td>\n<td>0.8392582947</td>\n<td>0.85750</td>\n<td>0.8447308063</td>\n</tr>\n<tr>\n<td>4</td>\n<td>CynapX</td>\n<td>0.8280631364</td>\n<td>0.86000</td>\n<td>0.8376441955</td>\n</tr>\n<tr>\n<td>5</td>\n<td>TeneT</td>\n<td>0.8191707574</td>\n<td>0.86625</td>\n<td>0.8332945302</td>\n</tr>\n<tr>\n<td>6</td>\n<td>HackFleet</td>\n<td>0.8337843439</td>\n<td>0.81750</td>\n<td>0.8288990407</td>\n</tr>\n<tr>\n<td>7</td>\n<td>zero-shot</td>\n<td>0.8396552538</td>\n<td>0.79500</td>\n<td>0.8262586777</td>\n</tr>\n<tr>\n<td>8</td>\n<td>BitwiseMind</td>\n<td>0.8417791523</td>\n<td>0.71750</td>\n<td>0.8044954066</td>\n</tr>\n<tr>\n<td>9</td>\n<td>KUET_TensorBit</td>\n<td>0.8163695686</td>\n<td>0.76000</td>\n<td>0.7994586980</td>\n</tr>\n<tr>\n<td>10</td>\n<td>Team ReActive</td>\n<td>0.8096149727</td>\n<td>0.77500</td>\n<td>0.7992304809</td>\n</tr>\n<tr>\n<td>11</td>\n<td>Luck is all you need</td>\n<td>0.8286634954</td>\n<td>0.72500</td>\n<td>0.7975644468</td>\n</tr>\n<tr>\n<td>12</td>\n<td>six_seven</td>\n<td>0.8152170980</td>\n<td>0.74750</td>\n<td>0.7949019686</td>\n</tr>\n<tr>\n<td>13</td>\n<td>Fellowship of the Ring</td>\n<td>0.8257633951</td>\n<td>0.72000</td>\n<td>0.7940343766</td>\n</tr>\n<tr>\n<td>14</td>\n<td>ArektaGenericTeam</td>\n<td>0.8288094745</td>\n<td>0.71250</td>\n<td>0.7939166322</td>\n</tr>\n<tr>\n<td>15</td>\n<td>BLUe</td>\n<td>0.8126221893</td>\n<td>0.71750</td>\n<td>0.7840855325</td>\n</tr>\n<tr>\n<td>16</td>\n<td>823-OLT</td>\n<td>0.8166729622</td>\n<td>0.67000</td>\n<td>0.7726710735</td>\n</tr>\n<tr>\n<td>17</td>\n<td>Quasar</td>\n<td>0.8111411476</td>\n<td>0.66750</td>\n<td>0.7680488034</td>\n</tr>\n<tr>\n<td>18</td>\n<td>Envisage</td>\n<td>0.8301329154</td>\n<td>0.49250</td>\n<td>0.7288430408</td>\n</tr>\n</tbody>\n</table>\n<hr>\n<blockquote>\n  <p><strong>Score Definitions (per Official Rulebook)</strong></p>\n  <ul>\n  <li><strong>Net Online Score (70% of Final)</strong> — Composite score derived from the Kaggle-hosted competition and a separate hidden test set, computed as:\n  <code>Net Online Score = 0.7 × Model Performance Score + 0.3 × Inference Time Score</code><ul>\n  <li><em>Model Performance Score</em> is a weighted average of the Public Leaderboard (15%), Private Leaderboard (35%), and Hidden Test Set (50%) evaluation metrics.</li>\n  <li><em>Inference Time Score</em> penalises slower submissions; Faster inference yields a higher score.</li></ul></li>\n  <li><strong>Net Offline Score (30% of Final)</strong> — Jury evaluation score awarded during the offline presentation round. Each team was assessed by a panel of four judges across three criteria — Presentation &amp; Q/A (20%), Paper Quality (20%), and Novelty (60%). The four judge totals were summed and normalised to a 0–1 scale.</li>\n  <li><strong>Final Weighted Score</strong> — The overall ranking metric, combining online and offline evaluations:\n  <code>Final Weighted Score = 0.7 × Net Online Score + 0.3 × Net Offline Score</code></li>\n  </ul>\n  <p>All scores are normalised to a <strong>0–1</strong> scale. The weighting has been verified against the official scoring sheet for all 18 teams.</p>\n</blockquote>\n<hr>\n<h2>Evaluation &amp; Scoring Breakdown (Official Rulebook)</h2>\n<h3>Overall Formula</h3>\n<p><strong>Final Weighted Score = 70% × Net Online Score + 30% × Net Offline Score</strong></p>\n<hr>\n<h3>Net Online Score — 70% of Final</h3>\n<p><strong>Net Online Score = 70% × Model Performance Score + 30% × Inference Time Score</strong></p>\n<table>\n<thead>\n<tr>\n<th>Component</th>\n<th>Weight</th>\n<th>Source</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Public Leaderboard Score</td>\n<td>15%</td>\n<td>Kaggle public dataset evaluation</td>\n</tr>\n<tr>\n<td>Private Leaderboard Score</td>\n<td>35%</td>\n<td>Kaggle private dataset evaluation</td>\n</tr>\n<tr>\n<td>Hidden Test Set Score</td>\n<td>50%</td>\n<td>Separate hidden test set provided by organisers</td>\n</tr>\n</tbody>\n</table>\n<p><em>Inference Time Score</em> is derived from the average inference latency of each submission, normalised such that faster models score higher.</p>\n<hr>\n<h3>Net Offline Score — 30% of Final</h3>\n<p>Evaluated by a panel of <strong>4 judges</strong> per team. Each judge scored out of the following:</p>\n<table>\n<thead>\n<tr>\n<th>Criterion</th>\n<th>Max per Judge</th>\n<th>Weight</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Presentation &amp; Q/A</td>\n<td>20</td>\n<td>20%</td>\n</tr>\n<tr>\n<td>Paper Quality</td>\n<td>20</td>\n<td>20%</td>\n</tr>\n<tr>\n<td>Novelty</td>\n<td>60</td>\n<td>60%</td>\n</tr>\n</tbody>\n</table>\n<p>Total offline marks = sum of all four judges' scores (out of 400), normalised to 0–1.</p>",
  "messages": [
    {
      "id": 3418008,
      "postDate": "2026-03-06T20:31:04.333Z",
      "content": "<h1>DL Sprint 4.0 — Final Scores &amp; Evaluation Breakdown</h1>\n<p>Congratulations to all the teams who made it to the top 18 — and an even bigger congratulations to every single participant who gave it their best. Reaching the final round out of <strong>192 registered teams</strong>, while competing in two intense competitions simultaneously, is no small feat. Regardless of where you placed, the effort, learning, and perseverance you demonstrated throughout this sprint speak for themselves.</p>\n<p>Best of luck to all of you in your future endeavours. </p>\n<p>In the spirit of transparency, we are making the complete score breakdown public so that every team can see exactly how the final rankings were determined.</p>\n<hr>\n<table>\n<thead>\n<tr>\n<th>Rank</th>\n<th>Team</th>\n<th>Net Online Score</th>\n<th>Net Offline Score</th>\n<th>Final Weighted Score</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>1</td>\n<td>MaraKheyeGechi</td>\n<td>0.8687789541</td>\n<td>0.80500</td>\n<td>0.8496452679</td>\n</tr>\n<tr>\n<td>2</td>\n<td>Labyrinth</td>\n<td>0.8339024755</td>\n<td>0.87625</td>\n<td>0.8466067328</td>\n</tr>\n<tr>\n<td>3</td>\n<td>Villagers</td>\n<td>0.8392582947</td>\n<td>0.85750</td>\n<td>0.8447308063</td>\n</tr>\n<tr>\n<td>4</td>\n<td>CynapX</td>\n<td>0.8280631364</td>\n<td>0.86000</td>\n<td>0.8376441955</td>\n</tr>\n<tr>\n<td>5</td>\n<td>TeneT</td>\n<td>0.8191707574</td>\n<td>0.86625</td>\n<td>0.8332945302</td>\n</tr>\n<tr>\n<td>6</td>\n<td>HackFleet</td>\n<td>0.8337843439</td>\n<td>0.81750</td>\n<td>0.8288990407</td>\n</tr>\n<tr>\n<td>7</td>\n<td>zero-shot</td>\n<td>0.8396552538</td>\n<td>0.79500</td>\n<td>0.8262586777</td>\n</tr>\n<tr>\n<td>8</td>\n<td>BitwiseMind</td>\n<td>0.8417791523</td>\n<td>0.71750</td>\n<td>0.8044954066</td>\n</tr>\n<tr>\n<td>9</td>\n<td>KUET_TensorBit</td>\n<td>0.8163695686</td>\n<td>0.76000</td>\n<td>0.7994586980</td>\n</tr>\n<tr>\n<td>10</td>\n<td>Team ReActive</td>\n<td>0.8096149727</td>\n<td>0.77500</td>\n<td>0.7992304809</td>\n</tr>\n<tr>\n<td>11</td>\n<td>Luck is all you need</td>\n<td>0.8286634954</td>\n<td>0.72500</td>\n<td>0.7975644468</td>\n</tr>\n<tr>\n<td>12</td>\n<td>six_seven</td>\n<td>0.8152170980</td>\n<td>0.74750</td>\n<td>0.7949019686</td>\n</tr>\n<tr>\n<td>13</td>\n<td>Fellowship of the Ring</td>\n<td>0.8257633951</td>\n<td>0.72000</td>\n<td>0.7940343766</td>\n</tr>\n<tr>\n<td>14</td>\n<td>ArektaGenericTeam</td>\n<td>0.8288094745</td>\n<td>0.71250</td>\n<td>0.7939166322</td>\n</tr>\n<tr>\n<td>15</td>\n<td>BLUe</td>\n<td>0.8126221893</td>\n<td>0.71750</td>\n<td>0.7840855325</td>\n</tr>\n<tr>\n<td>16</td>\n<td>823-OLT</td>\n<td>0.8166729622</td>\n<td>0.67000</td>\n<td>0.7726710735</td>\n</tr>\n<tr>\n<td>17</td>\n<td>Quasar</td>\n<td>0.8111411476</td>\n<td>0.66750</td>\n<td>0.7680488034</td>\n</tr>\n<tr>\n<td>18</td>\n<td>Envisage</td>\n<td>0.8301329154</td>\n<td>0.49250</td>\n<td>0.7288430408</td>\n</tr>\n</tbody>\n</table>\n<hr>\n<blockquote>\n  <p><strong>Score Definitions (per Official Rulebook)</strong></p>\n  <ul>\n  <li><strong>Net Online Score (70% of Final)</strong> — Composite score derived from the Kaggle-hosted competition and a separate hidden test set, computed as:\n  <code>Net Online Score = 0.7 × Model Performance Score + 0.3 × Inference Time Score</code><ul>\n  <li><em>Model Performance Score</em> is a weighted average of the Public Leaderboard (15%), Private Leaderboard (35%), and Hidden Test Set (50%) evaluation metrics.</li>\n  <li><em>Inference Time Score</em> penalises slower submissions; Faster inference yields a higher score.</li></ul></li>\n  <li><strong>Net Offline Score (30% of Final)</strong> — Jury evaluation score awarded during the offline presentation round. Each team was assessed by a panel of four judges across three criteria — Presentation &amp; Q/A (20%), Paper Quality (20%), and Novelty (60%). The four judge totals were summed and normalised to a 0–1 scale.</li>\n  <li><strong>Final Weighted Score</strong> — The overall ranking metric, combining online and offline evaluations:\n  <code>Final Weighted Score = 0.7 × Net Online Score + 0.3 × Net Offline Score</code></li>\n  </ul>\n  <p>All scores are normalised to a <strong>0–1</strong> scale. The weighting has been verified against the official scoring sheet for all 18 teams.</p>\n</blockquote>\n<hr>\n<h2>Evaluation &amp; Scoring Breakdown (Official Rulebook)</h2>\n<h3>Overall Formula</h3>\n<p><strong>Final Weighted Score = 70% × Net Online Score + 30% × Net Offline Score</strong></p>\n<hr>\n<h3>Net Online Score — 70% of Final</h3>\n<p><strong>Net Online Score = 70% × Model Performance Score + 30% × Inference Time Score</strong></p>\n<table>\n<thead>\n<tr>\n<th>Component</th>\n<th>Weight</th>\n<th>Source</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Public Leaderboard Score</td>\n<td>15%</td>\n<td>Kaggle public dataset evaluation</td>\n</tr>\n<tr>\n<td>Private Leaderboard Score</td>\n<td>35%</td>\n<td>Kaggle private dataset evaluation</td>\n</tr>\n<tr>\n<td>Hidden Test Set Score</td>\n<td>50%</td>\n<td>Separate hidden test set provided by organisers</td>\n</tr>\n</tbody>\n</table>\n<p><em>Inference Time Score</em> is derived from the average inference latency of each submission, normalised such that faster models score higher.</p>\n<hr>\n<h3>Net Offline Score — 30% of Final</h3>\n<p>Evaluated by a panel of <strong>4 judges</strong> per team. Each judge scored out of the following:</p>\n<table>\n<thead>\n<tr>\n<th>Criterion</th>\n<th>Max per Judge</th>\n<th>Weight</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Presentation &amp; Q/A</td>\n<td>20</td>\n<td>20%</td>\n</tr>\n<tr>\n<td>Paper Quality</td>\n<td>20</td>\n<td>20%</td>\n</tr>\n<tr>\n<td>Novelty</td>\n<td>60</td>\n<td>60%</td>\n</tr>\n</tbody>\n</table>\n<p>Total offline marks = sum of all four judges' scores (out of 400), normalised to 0–1.</p>",
      "rawMarkdown": "# DL Sprint 4.0 — Final Scores & Evaluation Breakdown\n\nCongratulations to all the teams who made it to the top 18 — and an even bigger congratulations to every single participant who gave it their best. Reaching the final round out of **192 registered teams**, while competing in two intense competitions simultaneously, is no small feat. Regardless of where you placed, the effort, learning, and perseverance you demonstrated throughout this sprint speak for themselves.\n\nBest of luck to all of you in your future endeavours. \n\nIn the spirit of transparency, we are making the complete score breakdown public so that every team can see exactly how the final rankings were determined.\n\n---\n\n\n| Rank | Team                   | Net Online Score | Net Offline Score | Final Weighted Score |\n| ---- | ---------------------- | ---------------- | ----------------- | -------------------- |\n| 1    | MaraKheyeGechi         | 0.8687789541     | 0.80500           | 0.8496452679         |\n| 2    | Labyrinth              | 0.8339024755     | 0.87625           | 0.8466067328         |\n| 3    | Villagers              | 0.8392582947     | 0.85750           | 0.8447308063         |\n| 4    | CynapX                 | 0.8280631364     | 0.86000           | 0.8376441955         |\n| 5    | TeneT                  | 0.8191707574     | 0.86625           | 0.8332945302         |\n| 6    | HackFleet              | 0.8337843439     | 0.81750           | 0.8288990407         |\n| 7    | zero-shot              | 0.8396552538     | 0.79500           | 0.8262586777         |\n| 8    | BitwiseMind            | 0.8417791523     | 0.71750           | 0.8044954066         |\n| 9    | KUET_TensorBit         | 0.8163695686     | 0.76000           | 0.7994586980         |\n| 10   | Team ReActive          | 0.8096149727     | 0.77500           | 0.7992304809         |\n| 11   | Luck is all you need   | 0.8286634954     | 0.72500           | 0.7975644468         |\n| 12   | six_seven              | 0.8152170980     | 0.74750           | 0.7949019686         |\n| 13   | Fellowship of the Ring | 0.8257633951     | 0.72000           | 0.7940343766         |\n| 14   | ArektaGenericTeam      | 0.8288094745     | 0.71250           | 0.7939166322         |\n| 15   | BLUe                   | 0.8126221893     | 0.71750           | 0.7840855325         |\n| 16   | 823-OLT                | 0.8166729622     | 0.67000           | 0.7726710735         |\n| 17   | Quasar                 | 0.8111411476     | 0.66750           | 0.7680488034         |\n| 18   | Envisage               | 0.8301329154     | 0.49250           | 0.7288430408         |\n\n\n---\n\n> **Score Definitions (per Official Rulebook)**\n>\n> - **Net Online Score (70% of Final)** — Composite score derived from the Kaggle-hosted competition and a separate hidden test set, computed as:\n> `Net Online Score = 0.7 × Model Performance Score + 0.3 × Inference Time Score`\n>   - *Model Performance Score* is a weighted average of the Public Leaderboard (15%), Private Leaderboard (35%), and Hidden Test Set (50%) evaluation metrics.\n>   - *Inference Time Score* penalises slower submissions; Faster inference yields a higher score.\n> - **Net Offline Score (30% of Final)** — Jury evaluation score awarded during the offline presentation round. Each team was assessed by a panel of four judges across three criteria — Presentation & Q/A (20%), Paper Quality (20%), and Novelty (60%). The four judge totals were summed and normalised to a 0–1 scale.\n> - **Final Weighted Score** — The overall ranking metric, combining online and offline evaluations:\n> `Final Weighted Score = 0.7 × Net Online Score + 0.3 × Net Offline Score`\n>\n> All scores are normalised to a **0–1** scale. The weighting has been verified against the official scoring sheet for all 18 teams.\n\n---\n\n## Evaluation & Scoring Breakdown (Official Rulebook)\n\n### Overall Formula\n\n**Final Weighted Score = 70% × Net Online Score + 30% × Net Offline Score**\n\n---\n\n### Net Online Score — 70% of Final\n\n**Net Online Score = 70% × Model Performance Score + 30% × Inference Time Score**\n\n\n| Component                 | Weight | Source                                          |\n| ------------------------- | ------ | ----------------------------------------------- |\n| Public Leaderboard Score  | 15%    | Kaggle public dataset evaluation                |\n| Private Leaderboard Score | 35%    | Kaggle private dataset evaluation               |\n| Hidden Test Set Score     | 50%    | Separate hidden test set provided by organisers |\n\n\n*Inference Time Score* is derived from the average inference latency of each submission, normalised such that faster models score higher.\n\n---\n\n### Net Offline Score — 30% of Final\n\nEvaluated by a panel of **4 judges** per team. Each judge scored out of the following:\n\n\n| Criterion          | Max per Judge | Weight |\n| ------------------ | ------------- | ------ |\n| Presentation & Q/A | 20            | 20%    |\n| Paper Quality      | 20            | 20%    |\n| Novelty            | 60            | 60%    |\n\n\nTotal offline marks = sum of all four judges' scores (out of 400), normalised to 0–1."
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3418008": "# DL Sprint 4.0 — Final Scores & Evaluation Breakdown\n\nCongratulations to all the teams who made it to the top 18 — and an even bigger congratulations to every single participant who gave it their best. Reaching the final round out of **192 registered teams**, while competing in two intense competitions simultaneously, is no small feat. Regardless of where you placed, the effort, learning, and perseverance you demonstrated throughout this sprint speak for themselves.\n\nBest of luck to all of you in your future endeavours. \n\nIn the spirit of transparency, we are making the complete score breakdown public so that every team can see exactly how the final rankings were determined.\n\n---\n\n\n| Rank | Team                   | Net Online Score | Net Offline Score | Final Weighted Score |\n| ---- | ---------------------- | ---------------- | ----------------- | -------------------- |\n| 1    | MaraKheyeGechi         | 0.8687789541     | 0.80500           | 0.8496452679         |\n| 2    | Labyrinth              | 0.8339024755     | 0.87625           | 0.8466067328         |\n| 3    | Villagers              | 0.8392582947     | 0.85750           | 0.8447308063         |\n| 4    | CynapX                 | 0.8280631364     | 0.86000           | 0.8376441955         |\n| 5    | TeneT                  | 0.8191707574     | 0.86625           | 0.8332945302         |\n| 6    | HackFleet              | 0.8337843439     | 0.81750           | 0.8288990407         |\n| 7    | zero-shot              | 0.8396552538     | 0.79500           | 0.8262586777         |\n| 8    | BitwiseMind            | 0.8417791523     | 0.71750           | 0.8044954066         |\n| 9    | KUET_TensorBit         | 0.8163695686     | 0.76000           | 0.7994586980         |\n| 10   | Team ReActive          | 0.8096149727     | 0.77500           | 0.7992304809         |\n| 11   | Luck is all you need   | 0.8286634954     | 0.72500           | 0.7975644468         |\n| 12   | six_seven              | 0.8152170980     | 0.74750           | 0.7949019686         |\n| 13   | Fellowship of the Ring | 0.8257633951     | 0.72000           | 0.7940343766         |\n| 14   | ArektaGenericTeam      | 0.8288094745     | 0.71250           | 0.7939166322         |\n| 15   | BLUe                   | 0.8126221893     | 0.71750           | 0.7840855325         |\n| 16   | 823-OLT                | 0.8166729622     | 0.67000           | 0.7726710735         |\n| 17   | Quasar                 | 0.8111411476     | 0.66750           | 0.7680488034         |\n| 18   | Envisage               | 0.8301329154     | 0.49250           | 0.7288430408         |\n\n\n---\n\n> **Score Definitions (per Official Rulebook)**\n>\n> - **Net Online Score (70% of Final)** — Composite score derived from the Kaggle-hosted competition and a separate hidden test set, computed as:\n> `Net Online Score = 0.7 × Model Performance Score + 0.3 × Inference Time Score`\n>   - *Model Performance Score* is a weighted average of the Public Leaderboard (15%), Private Leaderboard (35%), and Hidden Test Set (50%) evaluation metrics.\n>   - *Inference Time Score* penalises slower submissions; Faster inference yields a higher score.\n> - **Net Offline Score (30% of Final)** — Jury evaluation score awarded during the offline presentation round. Each team was assessed by a panel of four judges across three criteria — Presentation & Q/A (20%), Paper Quality (20%), and Novelty (60%). The four judge totals were summed and normalised to a 0–1 scale.\n> - **Final Weighted Score** — The overall ranking metric, combining online and offline evaluations:\n> `Final Weighted Score = 0.7 × Net Online Score + 0.3 × Net Offline Score`\n>\n> All scores are normalised to a **0–1** scale. The weighting has been verified against the official scoring sheet for all 18 teams.\n\n---\n\n## Evaluation & Scoring Breakdown (Official Rulebook)\n\n### Overall Formula\n\n**Final Weighted Score = 70% × Net Online Score + 30% × Net Offline Score**\n\n---\n\n### Net Online Score — 70% of Final\n\n**Net Online Score = 70% × Model Performance Score + 30% × Inference Time Score**\n\n\n| Component                 | Weight | Source                                          |\n| ------------------------- | ------ | ----------------------------------------------- |\n| Public Leaderboard Score  | 15%    | Kaggle public dataset evaluation                |\n| Private Leaderboard Score | 35%    | Kaggle private dataset evaluation               |\n| Hidden Test Set Score     | 50%    | Separate hidden test set provided by organisers |\n\n\n*Inference Time Score* is derived from the average inference latency of each submission, normalised such that faster models score higher.\n\n---\n\n### Net Offline Score — 30% of Final\n\nEvaluated by a panel of **4 judges** per team. Each judge scored out of the following:\n\n\n| Criterion          | Max per Judge | Weight |\n| ------------------ | ------------- | ------ |\n| Presentation & Q/A | 20            | 20%    |\n| Paper Quality      | 20            | 20%    |\n| Novelty            | 60            | 60%    |\n\n\nTotal offline marks = sum of all four judges' scores (out of 400), normalised to 0–1."
  }
}