Upload folder using huggingface_hub

README.md

@@ -8,4 +8,4 @@ models:
 - open-nhe/Elysium-X-500-FR
 pinned: false
 ---
-Private showcase for Elysium X 500 FR (results only, no data rows).
+Results showcase for Elysium X 500 FR (public-data upgrade). Results only, no data rows. CC BY-NC-SA 4.0.

index.html

@@ -1,9 +1,18 @@
-<!doctype html><meta charset=utf-8><title>Elysium X 500 FR</title><style>
+<!doctype html><meta charset=utf-8><title>Elysium X 500 FR</title><meta name=viewport content="width=device-width,initial-scale=1"><style>
 body{font-family:Inter,Segoe UI,Arial,sans-serif;max-width:900px;margin:auto;padding:16px;background:#0f172a;color:#e2e8f0}
+h1,h2,h3{background:linear-gradient(90deg,#f472b6,#a78bfa,#38bdf8);-webkit-background-clip:text;color:transparent}
 table{border-collapse:collapse;width:100%}td,th{border:1px solid #334155;padding:8px;text-align:left}
-.r{display:flex;align-items:center;gap:8px;margin:4px 0}.r span{width:70px}.r i{display:block;height:14px;background:#6366f1;border-radius:7px;animation:g 1.5s ease-out}
-@keyframes g{from{width:0}}.w{color:#fde68a}</style>
-<img src=assets/banner.svg style=width:100%><h2>Headline: 19.4% exact on rows without the emotion name (n=403)</h2>
-<table><tr><th>Test slice</th><th>Exact</th><th>Top-1</th><th>id-F1</th></tr><tr><td>No emotion name in text (n=403), headline</td><td>19.35%</td><td>20.84%</td><td>0.206</td></tr><tr><td>Emotion name in text (n=792)</td><td>95.08%</td><td>96.59%</td><td>0.971</td></tr><tr><td>All rows (n=1,195), not the headline</td><td>69.54%</td><td>71.05%</td><td>0.713</td></tr></table>
-<h3>No-name exact by language</h3><div class=r><span>PT</span><i style='width:81.8%'></i><b>40.9%</b></div><div class=r><span>FR</span><i style='width:64.2%'></i><b>32.1%</b></div><div class=r><span>ZH</span><i style='width:61.6%'></i><b>30.8%</b></div><div class=r><span>DE</span><i style='width:60.0%'></i><b>30.0%</b></div><div class=r><span>Hinglish</span><i style='width:58.4%'></i><b>29.2%</b></div><div class=r><span>EN</span><i style='width:52.6%'></i><b>26.3%</b></div><div class=r><span>JA</span><i style='width:34.4%'></i><b>17.2%</b></div><div class=r><span>AR</span><i style='width:31.6%'></i><b>15.8%</b></div><div class=r><span>ES</span><i style='width:29.6%'></i><b>14.8%</b></div><div class=r><span>KO</span><i style='width:27.8%'></i><b>13.9%</b></div><div class=r><span>BN</span><i style='width:14.2%'></i><b>7.1%</b></div><div class=r><span>HI</span><i style='width:7.6%'></i><b>3.8%</b></div>
-<p class=w>Caveats: test covers 368 of 500 coordinates; template-generated data; possible cross-language leakage; self-reported; dataset by the owner and his OpenNHE team, human-monitored, cleaning steps disclosed. Private; all rights reserved.</p>
+.r{display:flex;align-items:center;gap:8px;margin:4px 0}.r span{width:70px}.r i{display:block;height:14px;background:linear-gradient(90deg,#6366f1,#ec4899);border-radius:7px;animation:g 1.5s ease-out}
+@keyframes g{from{width:0}}.w{color:#fde68a}.up{color:#4ade80}.dn{color:#fb7185}a{color:#7dd3fc}</style>
+<img src=assets/banner.svg style=width:100%>
+<h1>Elysium X 500 FR: public-data upgrade</h1>
+<p>Same 500-emotion schema, weights continued-trained on the team dataset plus 8,250 mapped rows from 11 public emotion datasets.</p>
+<h2>Headline: 23.3% exact on rows without the emotion name (n=403), up from 19.4%</h2>
+<table><tr><th>Test slice (n=1,195)</th><th>Original</th><th>This release</th><th>Change</th></tr>
+<tr><td>No emotion name (n=403), headline</td><td>19.35%</td><td>23.33%</td><td class=up>+3.98</td></tr>
+<tr><td>All rows</td><td>69.54%</td><td>70.71%</td><td class=up>+1.17</td></tr>
+<tr><td>Emotion name in text (n=792)</td><td>95.08%</td><td>94.82%</td><td class=dn>-0.26 (2 of 792)</td></tr></table>
+<p class=w>We set a bar to beat all three numbers. This release beats two and misses the with-name number by 2 answers. We say so. A repeat run (with a training-loop bug) scored lower and is not released.</p>
+<h3>No-name exact by language</h3><div class=r><span>DE</span><i style='width:100.0%'></i><b>50.0%</b></div><div class=r><span>PT</span><i style='width:90.9%'></i><b>45.5%</b></div><div class=r><span>EN</span><i style='width:78.9%'></i><b>39.5%</b></div><div class=r><span>FR</span><i style='width:71.4%'></i><b>35.7%</b></div><div class=r><span>KO</span><i style='width:55.6%'></i><b>27.8%</b></div><div class=r><span>Hinglish</span><i style='width:50.0%'></i><b>25.0%</b></div><div class=r><span>ZH</span><i style='width:46.2%'></i><b>23.1%</b></div><div class=r><span>AR</span><i style='width:42.1%'></i><b>21.1%</b></div><div class=r><span>JA</span><i style='width:34.5%'></i><b>17.2%</b></div><div class=r><span>ES</span><i style='width:32.8%'></i><b>16.4%</b></div><div class=r><span>BN</span><i style='width:14.3%'></i><b>7.1%</b></div><div class=r><span>HI</span><i style='width:7.5%'></i><b>3.8%</b></div>
+<p class=w>Caveats: test covers part of the 500 coordinates; team data is synthetic and templated; labels are human-monitored, not gold; small cells per language; one run, no confidence intervals. Licence CC BY-NC-SA 4.0, non-commercial; some source datasets have unknown licences (listed on the model card). Not for medical or profiling use.</p>
+<p><a href="https://huggingface.co/open-nhe/Elysium-X-500-FR">Model</a> · <a href="https://huggingface.co/open-nhe/Elysium-X-500-FR-GGUF">GGUF</a> · <a href="https://huggingface.co/datasets/open-nhe/elysium-x-500-fr-dataset">Dataset</a> · <a href="https://huggingface.co/spaces/open-nhe/Elysium-X-500-FR-Paper">Paper</a></p>