| 1 | <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1280 720" width="1280" height="720"> |
| 2 | <g id="bg"> |
| 3 | <rect width="1280" height="720" fill="#FFFFFF" /> |
| 4 | </g> |
| 5 | |
| 6 | <g id="header"> |
| 7 | <rect x="60" y="56" width="6" height="40" fill="#3182CE" /> |
| 8 | <text x="84" y="88" font-family="Georgia, "Microsoft YaHei", serif" font-size="36" font-weight="700" fill="#1A365D">New SOTA on WMT'14 — at a fraction of the cost</text> |
| 9 | <text x="84" y="118" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="700" fill="#4A5568" letter-spacing="1.5">TABLE 2 · BLEU SCORES + TRAINING-COST FLOPS · BEST PRIOR ROWS VS TRANSFORMER</text> |
| 10 | </g> |
| 11 | |
| 12 | <g id="table"> |
| 13 | <rect x="60" y="160" width="1160" height="380" fill="none" stroke="#1A365D" stroke-width="2" /> |
| 14 | <g id="table-area" transform="translate(60, 160)"> |
| 15 | <rect x="0" y="0" width="1160" height="60" fill="#1A365D" /> |
| 16 | <line x1="0" y1="60" x2="1160" y2="60" stroke="#1A365D" stroke-width="2" /> |
| 17 | <text x="24" y="37" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="800" fill="#FFFFFF" letter-spacing="1.5">MODEL</text> |
| 18 | <text x="640" y="37" text-anchor="middle" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="800" fill="#FFFFFF" letter-spacing="1.5">EN → DE BLEU</text> |
| 19 | <text x="840" y="37" text-anchor="middle" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="800" fill="#FFFFFF" letter-spacing="1.5">EN → FR BLEU</text> |
| 20 | <text x="1040" y="37" text-anchor="middle" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="800" fill="#FFFFFF" letter-spacing="1.5">TRAIN FLOPs</text> |
| 21 | |
| 22 | <line x1="540" y1="0" x2="540" y2="380" stroke="#E2E8F0" stroke-width="1.5" /> |
| 23 | <line x1="740" y1="0" x2="740" y2="380" stroke="#E2E8F0" stroke-width="1.5" /> |
| 24 | <line x1="940" y1="0" x2="940" y2="380" stroke="#E2E8F0" stroke-width="1.5" /> |
| 25 | |
| 26 | <g id="row-bytenet"> |
| 27 | <line x1="0" y1="106" x2="1160" y2="106" stroke="#E2E8F0" stroke-width="1" /> |
| 28 | <text x="24" y="92" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="15" font-weight="600" fill="#1A202C">ByteNet</text> |
| 29 | <text x="640" y="92" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="16" fill="#1A202C">23.75</text> |
| 30 | <text x="840" y="92" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="16" fill="#4A5568">—</text> |
| 31 | <text x="1040" y="92" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="13" fill="#4A5568">—</text> |
| 32 | </g> |
| 33 | |
| 34 | <g id="row-conv-s2s"> |
| 35 | <line x1="0" y1="152" x2="1160" y2="152" stroke="#E2E8F0" stroke-width="1" /> |
| 36 | <text x="24" y="138" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="15" font-weight="600" fill="#1A202C">ConvS2S</text> |
| 37 | <text x="640" y="138" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="16" fill="#1A202C">25.16</text> |
| 38 | <text x="840" y="138" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="16" fill="#1A202C">40.46</text> |
| 39 | <text x="1040" y="138" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="13" fill="#4A5568">9.6 × 10¹⁸</text> |
| 40 | </g> |
| 41 | |
| 42 | <g id="row-moe"> |
| 43 | <line x1="0" y1="198" x2="1160" y2="198" stroke="#E2E8F0" stroke-width="1" /> |
| 44 | <text x="24" y="184" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="15" font-weight="600" fill="#1A202C">MoE</text> |
| 45 | <text x="640" y="184" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="16" fill="#1A202C">26.03</text> |
| 46 | <text x="840" y="184" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="16" fill="#1A202C">40.56</text> |
| 47 | <text x="1040" y="184" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="13" fill="#4A5568">2.0 × 10¹⁹</text> |
| 48 | </g> |
| 49 | |
| 50 | <g id="row-conv-s2s-ens"> |
| 51 | <line x1="0" y1="244" x2="1160" y2="244" stroke="#E2E8F0" stroke-width="1" /> |
| 52 | <text x="24" y="230" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="15" font-weight="600" fill="#1A202C">ConvS2S Ensemble</text> |
| 53 | <text x="640" y="230" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="16" fill="#1A202C">26.36</text> |
| 54 | <text x="840" y="230" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="16" fill="#1A202C">41.29</text> |
| 55 | <text x="1040" y="230" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="13" fill="#4A5568">7.7 × 10¹⁹</text> |
| 56 | </g> |
| 57 | |
| 58 | <g id="row-transformer-base"> |
| 59 | <rect x="0" y="244" width="1160" height="68" fill="#EBF4FB" /> |
| 60 | <line x1="0" y1="312" x2="1160" y2="312" stroke="#E2E8F0" stroke-width="1" /> |
| 61 | <text x="24" y="278" font-family="Georgia, "Microsoft YaHei", serif" font-size="17" font-weight="700" fill="#1A365D">Transformer (base)</text> |
| 62 | <text x="24" y="298" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="12" font-style="italic" fill="#3182CE">N=6, d_model=512, h=8</text> |
| 63 | <text x="640" y="284" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="18" font-weight="700" fill="#1A365D">27.3</text> |
| 64 | <text x="840" y="284" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="18" font-weight="700" fill="#1A365D">38.1</text> |
| 65 | <text x="1040" y="284" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="14" fill="#3182CE">3.3 × 10¹⁸</text> |
| 66 | </g> |
| 67 | |
| 68 | <g id="row-transformer-big"> |
| 69 | <rect x="0" y="312" width="1160" height="68" fill="#1A365D" fill-opacity="0.06" /> |
| 70 | <text x="24" y="346" font-family="Georgia, "Microsoft YaHei", serif" font-size="17" font-weight="700" fill="#1A365D">Transformer (big)</text> |
| 71 | <text x="24" y="366" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="12" font-style="italic" fill="#3182CE">N=6, d_model=1024, h=16</text> |
| 72 | <text x="640" y="352" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="22" font-weight="800" fill="#2F855A">28.4</text> |
| 73 | <text x="840" y="352" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="22" font-weight="800" fill="#2F855A">41.8</text> |
| 74 | <text x="1040" y="352" text-anchor="middle" font-family="Consolas, "Courier New", monospace" font-size="14" fill="#3182CE">2.3 × 10¹⁹</text> |
| 75 | </g> |
| 76 | </g> |
| 77 | </g> |
| 78 | |
| 79 | <g id="takeaway"> |
| 80 | <path fill="#F5F7FA" stroke="#E2E8F0" stroke-width="1" d="M70,562 H1210 A10,10 0 0 1 1220,572 V662 A10,10 0 0 1 1210,672 H70 A10,10 0 0 1 60,662 V572 A10,10 0 0 1 70,562 Z" /> |
| 81 | <path fill="#2F855A" d="M63,562 H63 A3,3 0 0 1 66,565 V669 A3,3 0 0 1 63,672 H63 A3,3 0 0 1 60,669 V565 A3,3 0 0 1 63,562 Z" /> |
| 82 | <text x="100" y="590" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="700" fill="#2F855A" letter-spacing="1.5">THREE RESULTS WORTH NAMING</text> |
| 83 | <text x="100" y="620" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="15" fill="#1A202C">· <tspan font-weight="700" fill="#1A365D">EN-DE</tspan> — Transformer (big) beats the prior best ensemble by <tspan fill="#2F855A" font-weight="700">+2.0 BLEU</tspan>, single model.</text> |
| 84 | <text x="100" y="644" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="15" fill="#1A202C">· <tspan font-weight="700" fill="#1A365D">EN-FR</tspan> — new <tspan fill="#2F855A" font-weight="700">single-model SOTA</tspan> at 41.8 BLEU.</text> |
| 85 | <text x="100" y="668" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="15" fill="#1A202C">· <tspan font-weight="700" fill="#1A365D">Training cost</tspan> — well below most strong baselines; base model trains in 12 hours.</text> |
| 86 | </g> |
| 87 | |
| 88 | <g id="footer"> |
| 89 | <text x="60" y="702" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="11" fill="#A0AEC0">Source: Vaswani et al. 2017, Table 2 (§6.1 Machine Translation)</text> |
| 90 | <text x="1220" y="702" text-anchor="end" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="11" fill="#A0AEC0">14</text> |
| 91 | </g> |
| 92 | </svg> |