Upload folder using huggingface_hub

README.md

@@ -7,7 +7,7 @@ sdk: static
 app_file: index.html
 pinned: false
 license: mit
-short_description: Jev 1.13 vs OpenJev vs NVIDIA in English RAG
+short_description: Jev 1.13 vs NVIDIA in English RAG, free to rerun
 ---
 
 Static leaderboard generated by `jev-rag publish` from real benchmark runs.

index.html

@@ -24,14 +24,14 @@
 <h1>Jev RAG Benchmark</h1>
 <p class="note">Free English benchmark of an open System One model as a reranking and decision layer in RAG.</p>
 <h2>Reranking (frozen top-20 candidates)</h2>
-<table><thead><tr><th>Dataset</th><th>n</th><th>Method</th><th>nDCG@10</th><th>Recall@5</th><th>MRR@10</th><th>Rerank p50</th></tr></thead><tbody><tr><td>scifact</td><td>300</td><td>A — no reranker (hybrid order)</td><td>71.67%</td><td>79.67%</td><td>68.75%</td><td>0 ms</td></tr><tr><td>scifact</td><td>300</td><td>J — OpenJev 0.1 batch noul (Codiv)</td><td>63.92%</td><td>68.67%</td><td>62.59%</td><td>1220 ms</td></tr><tr><td>scifact</td><td>300</td><td>N — NVIDIA reranker</td><td>78.70%</td><td>87.33%</td><td>76.16%</td><td>307 ms</td></tr><tr><td>scifact</td><td>300</td><td>T — TypeSafe Jev 1.13 batch noul (real Jev)</td><td>79.29%</td><td>85.67%</td><td>77.03%</td><td>4044 ms</td></tr><tr><td>xquad-en</td><td>1190</td><td>A — no reranker (hybrid order)</td><td>98.11%</td><td>99.58%</td><td>97.60%</td><td>0 ms</td></tr><tr><td>xquad-en</td><td>1190</td><td>J — OpenJev 0.1 batch noul (Codiv)</td><td>98.03%</td><td>98.82%</td><td>97.65%</td><td>923 ms</td></tr><tr><td>xquad-en</td><td>1190</td><td>N — NVIDIA reranker</td><td>99.37%</td><td>99.66%</td><td>99.27%</td><td>409 ms</td></tr><tr><td>xquad-en</td><td>1190</td><td>T — TypeSafe Jev 1.13 batch noul (real Jev)</td><td>98.93%</td><td>99.66%</td><td>98.67%</td><td>4000 ms</td></tr></tbody></table>
-<h2>OpenJev calibration (candidate-level relevance probabilities)</h2>
-<table><thead><tr><th>Dataset</th><th>ECE (10 bin)</th><th>Brier</th><th>Top-1 accuracy</th><th>Top-1 confidence (correct)</th><th>Top-1 confidence (wrong)</th></tr></thead><tbody><tr><td>scifact</td><td>0.0906</td><td>0.0521</td><td>57.67%</td><td>0.606</td><td>0.514</td></tr><tr><td>xquad-en</td><td>0.0245</td><td>0.0086</td><td>96.72%</td><td>0.765</td><td>0.455</td></tr></tbody></table>
-<h2>Frozen-context answer generation (OpenJev top-5)</h2>
-<table><thead><tr><th>Dataset</th><th>Generator</th><th>Token F1</th><th>Exact match</th><th>F1 &gt;= 0.5</th><th>Abstention</th><th>Valid citations</th></tr></thead><tbody><tr><td>xquad-en</td><td>diffusiongemma-26b</td><td>31.89%</td><td>2.35%</td><td>19.16%</td><td>2.77%</td><td>99.85%</td></tr></tbody></table>
-<h2>RAG optimization mode (confidence-partitioned OpenJev, fixed t = 0.50)</h2>
-<table><thead><tr><th>Dataset</th><th>Threshold</th><th>A baseline nDCG@10</th><th>J always-on nDCG@10</th><th>G partitioned nDCG@10</th><th>Delta vs baseline</th><th>95% CI</th></tr></thead><tbody><tr><td>scifact</td><td>0.50</td><td>71.67%</td><td>63.92%</td><td>71.49%</td><td>-0.18 pts</td><td>-1.93 to +1.54</td></tr><tr><td>xquad-en</td><td>0.50</td><td>98.11%</td><td>98.03%</td><td>98.76%</td><td>+0.66 pts</td><td>+0.27 to +1.07</td></tr></tbody></table>
-<footer>Generated 2026-09-20 17:35 UTC. Real runs only; fixture runs are never published. Code and raw results: <a href="https://github.com/emretheus/jev-rag-benchmark">https://github.com/emretheus/jev-rag-benchmark</a>. OpenJev is an independent open-weights model served by Codiv; not affiliated with TypeSafe AI.</footer>
+<table><thead><tr><th>Dataset</th><th>n</th><th>Method</th><th>nDCG@10</th><th>Recall@5</th><th>MRR@10</th><th>Rerank p50</th></tr></thead><tbody><tr><td>scifact</td><td>300</td><td>A — no reranker (hybrid order)</td><td>71.67%</td><td>79.67%</td><td>68.75%</td><td>0 ms</td></tr><tr><td>scifact</td><td>300</td><td>T — TypeSafe Jev 1.13 batch noul</td><td>79.29%</td><td>85.67%</td><td>77.03%</td><td>4044 ms</td></tr><tr><td>scifact</td><td>300</td><td>N — NVIDIA cross-encoder reranker</td><td>78.70%</td><td>87.33%</td><td>76.16%</td><td>307 ms</td></tr><tr><td>xquad-en</td><td>1190</td><td>A — no reranker (hybrid order)</td><td>98.11%</td><td>99.58%</td><td>97.60%</td><td>0 ms</td></tr><tr><td>xquad-en</td><td>1190</td><td>T — TypeSafe Jev 1.13 batch noul</td><td>98.93%</td><td>99.66%</td><td>98.67%</td><td>4000 ms</td></tr><tr><td>xquad-en</td><td>1190</td><td>N — NVIDIA cross-encoder reranker</td><td>99.37%</td><td>99.66%</td><td>99.27%</td><td>409 ms</td></tr></tbody></table>
+<h2>Probability calibration (candidate-level relevance)</h2>
+<table><thead><tr><th>Dataset</th><th>Model</th><th>ECE (10 bin)</th><th>Brier</th><th>Top-1 accuracy</th><th>Top-1 confidence (correct)</th><th>Top-1 confidence (wrong)</th></tr></thead><tbody><tr><td>scifact</td><td>TypeSafe Jev 1.13 batch noul</td><td>0.0625</td><td>0.0366</td><td>71.00%</td><td>0.812</td><td>0.549</td></tr><tr><td>xquad-en</td><td>TypeSafe Jev 1.13 batch noul</td><td>0.0133</td><td>0.0045</td><td>97.90%</td><td>0.960</td><td>0.740</td></tr></tbody></table>
+<h2>Frozen-context answer generation</h2>
+<table><thead><tr><th>Dataset</th><th>Generator</th><th>Token F1</th><th>Exact match</th><th>F1 &gt;= 0.5</th><th>Abstention</th><th>Valid citations</th></tr></thead><tbody><tr><td>xquad-en</td><td>T</td><td>diffusiongemma-26b</td><td>30.58%</td><td>1.26%</td><td>17.31%</td><td>2.61%</td><td>99.85%</td></tr></tbody></table>
+<h2>RAG optimization mode (confidence-partitioned Jev, fixed t = 0.50)</h2>
+<table><thead><tr><th>Dataset</th><th>Model</th><th>Threshold</th><th>A baseline nDCG@10</th><th>Always-on nDCG@10</th><th>Partitioned nDCG@10</th><th>Delta vs baseline</th><th>95% CI</th></tr></thead><tbody><tr><td>scifact</td><td>TypeSafe Jev 1.13 batch noul</td><td>0.50</td><td>71.67%</td><td>79.29%</td><td>75.81%</td><td>+4.14 pts</td><td>+2.13 to +6.29</td></tr><tr><td>xquad-en</td><td>TypeSafe Jev 1.13 batch noul</td><td>0.50</td><td>98.11%</td><td>98.93%</td><td>99.00%</td><td>+0.89 pts</td><td>+0.45 to +1.36</td></tr></tbody></table>
+<footer>Generated 2026-09-23 09:29 UTC. Real runs only; fixture runs are never published. Code and raw results: <a href="https://github.com/emretheus/jev-rag-benchmark">https://github.com/emretheus/jev-rag-benchmark</a>. Independent benchmark, not affiliated with TypeSafe AI. All runs on free tiers.</footer>
 </main>
 </body>
 </html>