ginigen-ai commited on
Commit
8635bfc
·
verified ·
1 Parent(s): 68d733b

엣지 모델 섹션 신설 — ginigen-ai/Edge-4B-TELL + GINIGEN TELL 개념·판별력 · 저장소 링크 신명칭 반영

Browse files
Files changed (4) hide show
  1. company.html +19 -0
  2. heligo.html +19 -0
  3. index.html +138 -2
  4. open.html +20 -1
company.html CHANGED
@@ -123,6 +123,25 @@
123
  .fig span{display:block; font-size:13px; color:var(--ink2); margin-top:5px}
124
  .figcap{font-size:12.5px; color:var(--ink3); margin:32px 0 0}
125
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
126
  .cards{display:grid; grid-template-columns:repeat(auto-fit,minmax(250px,1fr));
127
  gap:18px; margin-top:52px; text-align:left}
128
  .card{background:var(--paper); border-radius:18px; padding:28px; box-shadow:0 2px 18px rgba(0,0,0,.06)}
 
123
  .fig span{display:block; font-size:13px; color:var(--ink2); margin-top:5px}
124
  .figcap{font-size:12.5px; color:var(--ink3); margin:32px 0 0}
125
 
126
+ /* 엣지 모델 섹션 */
127
+ .mdl{display:grid; grid-template-columns:1fr 1fr; gap:34px; align-items:start; margin-top:34px}
128
+ .mdl .pane{background:var(--paper); border-radius:18px; padding:28px;
129
+ box-shadow:0 2px 18px rgba(0,0,0,.06)}
130
+ .mdl .pane.edge{border-top:3px solid var(--accent)}
131
+ .repo{display:inline-flex; align-items:center; gap:9px; font-weight:600; font-size:15px;
132
+ padding:9px 15px; border-radius:10px; background:var(--soft); color:var(--accent-ink);
133
+ text-decoration:none; margin-top:4px}
134
+ .repo:hover{background:#E8EFEE; text-decoration:none}
135
+ .gauge{display:grid; gap:9px; margin:16px 0 4px}
136
+ .gauge div{display:grid; grid-template-columns:1fr auto; gap:12px; align-items:center;
137
+ font-size:14px}
138
+ .gauge i{display:block; height:7px; border-radius:99px; background:var(--rule);
139
+ grid-column:1/-1; overflow:hidden}
140
+ .gauge i em{display:block; height:100%; background:var(--ink3); border-radius:99px}
141
+ .gauge .win i em{background:var(--accent)}
142
+ .gauge b{font-variant-numeric:tabular-nums; font-weight:600}
143
+ .gauge .dim{color:var(--ink3)}
144
+ @media (max-width:820px){ .mdl{grid-template-columns:1fr; gap:22px} }
145
  .cards{display:grid; grid-template-columns:repeat(auto-fit,minmax(250px,1fr));
146
  gap:18px; margin-top:52px; text-align:left}
147
  .card{background:var(--paper); border-radius:18px; padding:28px; box-shadow:0 2px 18px rgba(0,0,0,.06)}
heligo.html CHANGED
@@ -123,6 +123,25 @@
123
  .fig span{display:block; font-size:13px; color:var(--ink2); margin-top:5px}
124
  .figcap{font-size:12.5px; color:var(--ink3); margin:32px 0 0}
125
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
126
  .cards{display:grid; grid-template-columns:repeat(auto-fit,minmax(250px,1fr));
127
  gap:18px; margin-top:52px; text-align:left}
128
  .card{background:var(--paper); border-radius:18px; padding:28px; box-shadow:0 2px 18px rgba(0,0,0,.06)}
 
123
  .fig span{display:block; font-size:13px; color:var(--ink2); margin-top:5px}
124
  .figcap{font-size:12.5px; color:var(--ink3); margin:32px 0 0}
125
 
126
+ /* 엣지 모델 섹션 */
127
+ .mdl{display:grid; grid-template-columns:1fr 1fr; gap:34px; align-items:start; margin-top:34px}
128
+ .mdl .pane{background:var(--paper); border-radius:18px; padding:28px;
129
+ box-shadow:0 2px 18px rgba(0,0,0,.06)}
130
+ .mdl .pane.edge{border-top:3px solid var(--accent)}
131
+ .repo{display:inline-flex; align-items:center; gap:9px; font-weight:600; font-size:15px;
132
+ padding:9px 15px; border-radius:10px; background:var(--soft); color:var(--accent-ink);
133
+ text-decoration:none; margin-top:4px}
134
+ .repo:hover{background:#E8EFEE; text-decoration:none}
135
+ .gauge{display:grid; gap:9px; margin:16px 0 4px}
136
+ .gauge div{display:grid; grid-template-columns:1fr auto; gap:12px; align-items:center;
137
+ font-size:14px}
138
+ .gauge i{display:block; height:7px; border-radius:99px; background:var(--rule);
139
+ grid-column:1/-1; overflow:hidden}
140
+ .gauge i em{display:block; height:100%; background:var(--ink3); border-radius:99px}
141
+ .gauge .win i em{background:var(--accent)}
142
+ .gauge b{font-variant-numeric:tabular-nums; font-weight:600}
143
+ .gauge .dim{color:var(--ink3)}
144
+ @media (max-width:820px){ .mdl{grid-template-columns:1fr; gap:22px} }
145
  .cards{display:grid; grid-template-columns:repeat(auto-fit,minmax(250px,1fr));
146
  gap:18px; margin-top:52px; text-align:left}
147
  .card{background:var(--paper); border-radius:18px; padding:28px; box-shadow:0 2px 18px rgba(0,0,0,.06)}
index.html CHANGED
@@ -123,6 +123,25 @@
123
  .fig span{display:block; font-size:13px; color:var(--ink2); margin-top:5px}
124
  .figcap{font-size:12.5px; color:var(--ink3); margin:32px 0 0}
125
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
126
  .cards{display:grid; grid-template-columns:repeat(auto-fit,minmax(250px,1fr));
127
  gap:18px; margin-top:52px; text-align:left}
128
  .card{background:var(--paper); border-radius:18px; padding:28px; box-shadow:0 2px 18px rgba(0,0,0,.06)}
@@ -307,6 +326,123 @@
307
  <p class="figcap" data-en>All figures come from public leaderboards and measurements on real hardware.</p>
308
  </div>
309
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
310
  <!-- 기술적 특징 -->
311
  <section class="soft"><div class="wrap center">
312
  <h2 class="h-lg" data-ko>엣지AI를 실제로 굴리는 여섯 가지</h2>
@@ -321,13 +457,13 @@
321
  <p class="body" data-ko>
322
  폰 안에서 추론 서버를 띄웁니다. ARM 점곱·행렬 명령을 켜 빌드해
323
  같은 칩에서 더 빨리 돌고, 실행파일은 <strong>12.6MB</strong>입니다.
324
- 가중치도 <a href="https://huggingface.co/ginigen-ai/heligo-gemma4-e4b-q4_0" target="_blank" rel="noopener">우리 저장소</a>에서
325
  직접 배포합니다 — 재난 앱이 남의 저장소에 목숨을 걸 수는 없습니다.
326
  </p>
327
  <p class="body" data-en>
328
  We run an inference server inside the phone, built with ARM dot-product and matrix
329
  instructions enabled. The runtime binary is <strong>12.6 MB</strong>, and the weights are
330
- served from <a href="https://huggingface.co/ginigen-ai/heligo-gemma4-e4b-q4_0" target="_blank" rel="noopener">our own repository</a> —
331
  a disaster app cannot stake its life on someone else's.
332
  </p>
333
  </div>
 
123
  .fig span{display:block; font-size:13px; color:var(--ink2); margin-top:5px}
124
  .figcap{font-size:12.5px; color:var(--ink3); margin:32px 0 0}
125
 
126
+ /* 엣지 모델 섹션 */
127
+ .mdl{display:grid; grid-template-columns:1fr 1fr; gap:34px; align-items:start; margin-top:34px}
128
+ .mdl .pane{background:var(--paper); border-radius:18px; padding:28px;
129
+ box-shadow:0 2px 18px rgba(0,0,0,.06)}
130
+ .mdl .pane.edge{border-top:3px solid var(--accent)}
131
+ .repo{display:inline-flex; align-items:center; gap:9px; font-weight:600; font-size:15px;
132
+ padding:9px 15px; border-radius:10px; background:var(--soft); color:var(--accent-ink);
133
+ text-decoration:none; margin-top:4px}
134
+ .repo:hover{background:#E8EFEE; text-decoration:none}
135
+ .gauge{display:grid; gap:9px; margin:16px 0 4px}
136
+ .gauge div{display:grid; grid-template-columns:1fr auto; gap:12px; align-items:center;
137
+ font-size:14px}
138
+ .gauge i{display:block; height:7px; border-radius:99px; background:var(--rule);
139
+ grid-column:1/-1; overflow:hidden}
140
+ .gauge i em{display:block; height:100%; background:var(--ink3); border-radius:99px}
141
+ .gauge .win i em{background:var(--accent)}
142
+ .gauge b{font-variant-numeric:tabular-nums; font-weight:600}
143
+ .gauge .dim{color:var(--ink3)}
144
+ @media (max-width:820px){ .mdl{grid-template-columns:1fr; gap:22px} }
145
  .cards{display:grid; grid-template-columns:repeat(auto-fit,minmax(250px,1fr));
146
  gap:18px; margin-top:52px; text-align:left}
147
  .card{background:var(--paper); border-radius:18px; padding:28px; box-shadow:0 2px 18px rgba(0,0,0,.06)}
 
326
  <p class="figcap" data-en>All figures come from public leaderboards and measurements on real hardware.</p>
327
  </div>
328
 
329
+ <!-- 엣지 모델 -->
330
+ <section id="model"><div class="wrap center">
331
+ <p class="eyebrow" data-ko>엣지 모델</p><p class="eyebrow" data-en>The edge model</p>
332
+ <h2 class="h-lg" data-ko>헬리GO 안에서 실제로 도는 모델</h2>
333
+ <h2 class="h-lg" data-en>The model that actually runs inside HeliGO</h2>
334
+ <p class="tagline center" data-ko>내려받아 그대로 돌려 보실 수 있습니다.</p>
335
+ <p class="tagline center" data-en>You can download it and run the same thing.</p>
336
+ <p><a class="repo" href="https://huggingface.co/ginigen-ai/Edge-4B-TELL" target="_blank" rel="noopener">
337
+ 🤗 ginigen-ai/Edge-4B-TELL</a></p>
338
+
339
+ <div class="mdl" style="text-align:left">
340
+ <div class="pane edge">
341
+ <span class="n">GINIGEN Edge 4B</span>
342
+ <h3 class="h-md" data-ko>폰 한 대 안에서 끝납니다</h3>
343
+ <h3 class="h-md" data-en>It ends inside one phone</h3>
344
+ <p class="body" data-ko>
345
+ 갤럭시 S25 에서 잰 값입니다. 가중치는 구글 <strong>Gemma 4</strong> 의
346
+ 양자화 인지 학습(QAT) 체크포인트를 <strong>손대지 않고 그대로</strong> 싣습니다 —
347
+ 모델의 공은 구글 딥마인드에 있습니다. 다만 <strong>우리 저장소에서 직접 배포</strong>합니다.
348
+ 재난 앱이 남의 저장소 주소에 목숨을 걸 수는 없습니다.
349
+ </p>
350
+ <p class="body" data-en>
351
+ Measured on a Galaxy S25. The weights are Google's <strong>Gemma 4</strong>
352
+ quantization-aware checkpoint, carried <strong>unmodified</strong> — all model credit
353
+ belongs to Google DeepMind. We serve them from <strong>our own repository</strong>:
354
+ a disaster app cannot stake its life on a path someone else controls.
355
+ </p>
356
+ <div class="kv">
357
+ <div><span data-ko>추론에 필요한 통신</span><span data-en>Network for inference</span><b>0</b></div>
358
+ <div><span data-ko>모델 올린 뒤 상주 메모리</span><span data-en>Resident memory</span><b>~3.6 GB</b></div>
359
+ <div><span data-ko>APK 안 추론 실행파일</span><span data-en>Runtime in the APK</span><b>12.6 MB</b></div>
360
+ <div><span data-ko>가중치 · 영상/음성 투영기</span><span data-en>Weights · projector</span><b>4.8 + 0.92 GB</b></div>
361
+ </div>
362
+ </div>
363
+
364
+ <div class="pane">
365
+ <span class="n">GINIGEN TELL</span>
366
+ <h3 class="h-md" data-ko>확신을 묻지 않고, 읽습니다</h3>
367
+ <h3 class="h-md" data-en>We do not ask for confidence. We read it.</h3>
368
+ <p class="body" data-ko>
369
+ 모델에게 “얼마나 확신하느냐”고 물으면 <strong>거의 항상 만점을 부릅니다.</strong>
370
+ 한국어 재난 절차 40문항에서 스스로 신고한 확신은 평균 <strong>0.994</strong>였고,
371
+ <strong>틀린 답에도 그랬습니다.</strong> 바늘이 늘 가득 찬 계기는 계기가 아닙니다.
372
+ </p>
373
+ <p class="body" data-en>
374
+ Ask a model how sure it is and it says <strong>almost always certain.</strong> Across 40
375
+ Korean disaster-procedure questions, self-reported confidence averaged
376
+ <strong>0.994</strong> — <strong>including on its wrong answers.</strong> A gauge that
377
+ always reads full is not a gauge.
378
+ </p>
379
+ <p class="body" data-ko>
380
+ <strong>TELL</strong>은 묻는 대신 모델의 <strong>내부 상태를 한 번 읽습니다.</strong>
381
+ 포커에서 <em>tell</em>이 본인도 모르게 새어 나오는 신호인 것처럼, 모델도 틀리기 직전에
382
+ 흔들립니다. 그 흔들림은 <strong>말보다 먼저</strong> 나타납니다.
383
+ </p>
384
+ <p class="body" data-en>
385
+ <strong>TELL</strong> reads the model's <strong>internal state</strong> instead of asking.
386
+ In poker a <em>tell</em> is the signal a player leaks without meaning to; a model wobbles
387
+ the same way, and the wobble shows up <strong>before the words do.</strong>
388
+ </p>
389
+
390
+ <p class="body" style="margin-bottom:2px" data-ko><strong>맞은 답을 가려내는 힘</strong>
391
+ <span class="dim" style="color:var(--ink3)">— Gemma-4-12B · 한국어 재난 절차 227문항 · 교차검증</span></p>
392
+ <p class="body" style="margin-bottom:2px" data-en><strong>Power to tell right from wrong</strong>
393
+ <span style="color:var(--ink3)">— Gemma-4-12B · 227 Korean questions · cross-validated</span></p>
394
+ <div class="gauge">
395
+ <div><span data-ko>모델이 스스로 신고한 확신</span><span data-en>Self-reported confidence</span><b>0.50</b></div>
396
+ <i><em style="width:50%"></em></i>
397
+ <div><span data-ko>겉모습 단서 (길이·숫자 개수)</span><span data-en>Surface cues (length, digits)</span><b>0.614</b></div>
398
+ <i><em style="width:61.4%"></em></i>
399
+ <div class="win"><span data-ko><strong>GINIGEN TELL</strong></span><span data-en><strong>GINIGEN TELL</strong></span><b>0.732</b></div>
400
+ <i class="win"><em style="width:73.2%; background:var(--accent)"></em></i>
401
+ </div>
402
+ <p class="body" data-ko>
403
+ 0.50은 동전 던지기입니다. 겉모습 단서만으로도 0.614가 나오기 때문에,
404
+ <strong>그 선을 넘지 못하면 아무것도 읽은 것이 아닙니다.</strong> TELL은 그 위로
405
+ <strong>0.118</strong> 올라갑니다.
406
+ </p>
407
+ <p class="body" data-en>
408
+ 0.50 is a coin flip. Surface cues alone already reach 0.614, so
409
+ <strong>clearing that line is the bar that matters.</strong> TELL clears it by
410
+ <strong>0.118</strong>.
411
+ </p>
412
+ <p class="body" data-ko>
413
+ 읽는 데 <strong>생성 토큰이 한 개도 들지 않습니다.</strong> 모델이 이미 한 계산을
414
+ 들여다볼 뿐입니다. 배터리로 도는 기기에서 이 차이는 “판정을 붙이느냐 마느냐”를 가릅니다.
415
+ </p>
416
+ <p class="body" data-en>
417
+ Reading it costs <strong>zero generated tokens</strong> — it looks at a computation the
418
+ model already performed. On battery-powered hardware that is the difference between
419
+ having calibration and skipping it.
420
+ </p>
421
+ <p class="body" data-ko>
422
+ <strong>TELL은 정답을 말하지 않습니다.</strong> “이 답이 흔들렸는가”만 말합니다.
423
+ 낮게 나오면 답을 <strong>공공기관 원문으로 내려보냅니다.</strong> 답을 승인하는 데는
424
+ 쓰지 않습니다 — 아래 안전 관문이 그 이유입니다.
425
+ </p>
426
+ <p class="body" data-en>
427
+ <strong>TELL never says what the right answer is.</strong> It says whether this answer
428
+ wobbled. A low score routes the reply <strong>down to the official source text.</strong>
429
+ It is never used to approve an answer — the safety gate below explains why.
430
+ </p>
431
+ <p class="body" style="color:var(--ink3)" data-ko>
432
+ 판독기는 모델마다 새로 맞춰야 합니다. 체크포인트가 다르면 내부 상태의 모양 자체가
433
+ 달라서, 한 모델의 판독기를 다른 모델에 그대로 옮길 수 없습니다.
434
+ 이 체크포인트용 판독기는 같은 절차로 맞추고 있고, 측정치는 같은 기준선과 함께 공개합니다.
435
+ </p>
436
+ <p class="body" style="color:var(--ink3)" data-en>
437
+ A readout is specific to the model it was fitted on: different checkpoints expose
438
+ differently shaped internal states, so a readout does not carry across. The one for this
439
+ checkpoint is being fitted the same way, and its measurement will be published with the
440
+ same baselines.
441
+ </p>
442
+ </div>
443
+ </div>
444
+ </div></section>
445
+
446
  <!-- 기술적 특징 -->
447
  <section class="soft"><div class="wrap center">
448
  <h2 class="h-lg" data-ko>엣지AI를 실제로 굴리는 여섯 가지</h2>
 
457
  <p class="body" data-ko>
458
  폰 안에서 추론 서버를 띄웁니다. ARM 점곱·행렬 명령을 켜 빌드해
459
  같은 칩에서 더 빨리 돌고, 실행파일은 <strong>12.6MB</strong>입니다.
460
+ 가중치도 <a href="https://huggingface.co/ginigen-ai/Edge-4B-TELL" target="_blank" rel="noopener">우리 저장소</a>에서
461
  직접 배포합니다 — 재난 앱이 남의 저장소에 목숨을 걸 수는 없습니다.
462
  </p>
463
  <p class="body" data-en>
464
  We run an inference server inside the phone, built with ARM dot-product and matrix
465
  instructions enabled. The runtime binary is <strong>12.6 MB</strong>, and the weights are
466
+ served from <a href="https://huggingface.co/ginigen-ai/Edge-4B-TELL" target="_blank" rel="noopener">our own repository</a> —
467
  a disaster app cannot stake its life on someone else's.
468
  </p>
469
  </div>
open.html CHANGED
@@ -123,6 +123,25 @@
123
  .fig span{display:block; font-size:13px; color:var(--ink2); margin-top:5px}
124
  .figcap{font-size:12.5px; color:var(--ink3); margin:32px 0 0}
125
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
126
  .cards{display:grid; grid-template-columns:repeat(auto-fit,minmax(250px,1fr));
127
  gap:18px; margin-top:52px; text-align:left}
128
  .card{background:var(--paper); border-radius:18px; padding:28px; box-shadow:0 2px 18px rgba(0,0,0,.06)}
@@ -321,7 +340,7 @@
321
  <span data-en>2nd on K-AI · 0.549, 0.011 behind first · CLIcK 0.791 · Apache 2.0</span></li>
322
  <li><a href="https://huggingface.co/ginigen-ai/Rogue-28B-MIX" target="_blank" rel="noopener">Rogue-28B-MIX <i>♥10</i></a>
323
  <span data-ko>혼합 학습 계열</span><span data-en>Mixed-training line</span></li>
324
- <li><a href="https://huggingface.co/ginigen-ai/heligo-gemma4-e4b-q4_0" target="_blank" rel="noopener">heligo-gemma4-e4b-q4_0</a>
325
  <span data-ko>헬리GO 에 실제로 실리는 배포 가중치 · 우리가 직접 호스팅합니다</span>
326
  <span data-en>The exact weights shipped in HeliGO — hosted by us, not borrowed</span></li>
327
  </ul>
 
123
  .fig span{display:block; font-size:13px; color:var(--ink2); margin-top:5px}
124
  .figcap{font-size:12.5px; color:var(--ink3); margin:32px 0 0}
125
 
126
+ /* 엣지 모델 섹션 */
127
+ .mdl{display:grid; grid-template-columns:1fr 1fr; gap:34px; align-items:start; margin-top:34px}
128
+ .mdl .pane{background:var(--paper); border-radius:18px; padding:28px;
129
+ box-shadow:0 2px 18px rgba(0,0,0,.06)}
130
+ .mdl .pane.edge{border-top:3px solid var(--accent)}
131
+ .repo{display:inline-flex; align-items:center; gap:9px; font-weight:600; font-size:15px;
132
+ padding:9px 15px; border-radius:10px; background:var(--soft); color:var(--accent-ink);
133
+ text-decoration:none; margin-top:4px}
134
+ .repo:hover{background:#E8EFEE; text-decoration:none}
135
+ .gauge{display:grid; gap:9px; margin:16px 0 4px}
136
+ .gauge div{display:grid; grid-template-columns:1fr auto; gap:12px; align-items:center;
137
+ font-size:14px}
138
+ .gauge i{display:block; height:7px; border-radius:99px; background:var(--rule);
139
+ grid-column:1/-1; overflow:hidden}
140
+ .gauge i em{display:block; height:100%; background:var(--ink3); border-radius:99px}
141
+ .gauge .win i em{background:var(--accent)}
142
+ .gauge b{font-variant-numeric:tabular-nums; font-weight:600}
143
+ .gauge .dim{color:var(--ink3)}
144
+ @media (max-width:820px){ .mdl{grid-template-columns:1fr; gap:22px} }
145
  .cards{display:grid; grid-template-columns:repeat(auto-fit,minmax(250px,1fr));
146
  gap:18px; margin-top:52px; text-align:left}
147
  .card{background:var(--paper); border-radius:18px; padding:28px; box-shadow:0 2px 18px rgba(0,0,0,.06)}
 
340
  <span data-en>2nd on K-AI · 0.549, 0.011 behind first · CLIcK 0.791 · Apache 2.0</span></li>
341
  <li><a href="https://huggingface.co/ginigen-ai/Rogue-28B-MIX" target="_blank" rel="noopener">Rogue-28B-MIX <i>♥10</i></a>
342
  <span data-ko>혼합 학습 계열</span><span data-en>Mixed-training line</span></li>
343
+ <li><a href="https://huggingface.co/ginigen-ai/Edge-4B-TELL" target="_blank" rel="noopener">Edge-4B-TELL</a>
344
  <span data-ko>헬리GO 에 실제로 실리는 배포 가중치 · 우리가 직접 호스팅합니다</span>
345
  <span data-en>The exact weights shipped in HeliGO — hosted by us, not borrowed</span></li>
346
  </ul>