nalanda-data commited on
Commit
ac8fc69
·
verified ·
1 Parent(s): 21d6f51

Reorder Benchmark tab: DrishtiTable before NalandaBench

Browse files
Files changed (1) hide show
  1. index.html +13 -13
index.html CHANGED
@@ -142,19 +142,7 @@
142
 
143
  <!-- BOARD -->
144
  <section class="panel" id="board">
145
- <h2>NalandaBenchJEE/NEET MCQ accuracy</h2>
146
- <p class="desc">Held-out multiple-choice questions. GRPO / RLVR post-training vs the open base model.</p>
147
- <table class="lb">
148
- <thead><tr><th>Model</th><th>Method</th><th style="text-align:right">Accuracy</th></tr></thead>
149
- <tbody>
150
- <tr><td>Nalanda Qwen 7B</td><td>GRPO / RLVR</td><td class="n">66.8%</td></tr>
151
- <tr><td>Qwen2.5-7B-Instruct</td><td>Base (no RLVR)</td><td class="n">60.5%</td></tr>
152
- <tr><td colspan="2">Improvement from RLVR post-training</td><td class="n delta">+6.3 pp</td></tr>
153
- </tbody>
154
- </table>
155
- <p class="note">Evaluated on 800 held-out MCQs from NalandaJEENEETBench. Figures are indicative and are refreshed as the benchmark split is finalized.</p>
156
-
157
- <h2 style="margin-top:30px">DrishtiTable — Table Structure Recognition (TEDS)</h2>
158
  <p class="desc">Image → HTML table recognition on 135 held-out Indian-textbook tables. Ranked by TEDS (the field-standard TSR metric). Our fine-tuned 7B model beats every zero-shot frontier model.</p>
159
  <table class="lb">
160
  <thead><tr><th>Model</th><th>Method</th><th style="text-align:right">TEDS</th><th style="text-align:right">S-TEDS</th></tr></thead>
@@ -172,6 +160,18 @@
172
  <a href="https://huggingface.co/spaces/nalanda-data/DrishtiTable-Leaderboard" target="_blank" rel="noopener">live leaderboard</a> ·
173
  <a href="https://huggingface.co/datasets/Nalandadata/DrishtiTable" target="_blank" rel="noopener">benchmark dataset</a>.
174
  </p>
 
 
 
 
 
 
 
 
 
 
 
 
175
  </section>
176
 
177
  <footer>
 
142
 
143
  <!-- BOARD -->
144
  <section class="panel" id="board">
145
+ <h2>DrishtiTableTable Structure Recognition (TEDS)</h2>
 
 
 
 
 
 
 
 
 
 
 
 
146
  <p class="desc">Image → HTML table recognition on 135 held-out Indian-textbook tables. Ranked by TEDS (the field-standard TSR metric). Our fine-tuned 7B model beats every zero-shot frontier model.</p>
147
  <table class="lb">
148
  <thead><tr><th>Model</th><th>Method</th><th style="text-align:right">TEDS</th><th style="text-align:right">S-TEDS</th></tr></thead>
 
160
  <a href="https://huggingface.co/spaces/nalanda-data/DrishtiTable-Leaderboard" target="_blank" rel="noopener">live leaderboard</a> ·
161
  <a href="https://huggingface.co/datasets/Nalandadata/DrishtiTable" target="_blank" rel="noopener">benchmark dataset</a>.
162
  </p>
163
+
164
+ <h2 style="margin-top:30px">NalandaBench — JEE/NEET MCQ accuracy</h2>
165
+ <p class="desc">Held-out multiple-choice questions. GRPO / RLVR post-training vs the open base model.</p>
166
+ <table class="lb">
167
+ <thead><tr><th>Model</th><th>Method</th><th style="text-align:right">Accuracy</th></tr></thead>
168
+ <tbody>
169
+ <tr><td>Nalanda Qwen 7B</td><td>GRPO / RLVR</td><td class="n">66.8%</td></tr>
170
+ <tr><td>Qwen2.5-7B-Instruct</td><td>Base (no RLVR)</td><td class="n">60.5%</td></tr>
171
+ <tr><td colspan="2">Improvement from RLVR post-training</td><td class="n delta">+6.3 pp</td></tr>
172
+ </tbody>
173
+ </table>
174
+ <p class="note">Evaluated on 800 held-out MCQs from NalandaJEENEETBench. Figures are indicative and are refreshed as the benchmark split is finalized.</p>
175
  </section>
176
 
177
  <footer>