| <html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>Distributed DPO reproduction</title><style> | |
| body{margin:0;font-family:Inter,system-ui,sans-serif;background:#0b1020;color:#eef2ff}.wrap{padding:28px;max-width:1100px;margin:auto}h1{font-size:34px;margin:0 0 8px}.sub{color:#a5b4fc;margin-bottom:24px}.grid{display:grid;grid-template-columns:repeat(3,1fr);gap:16px}.card{background:#151c34;border:1px solid #303b67;border-radius:14px;padding:18px}.n{font-size:30px;font-weight:800;color:#67e8f9}.ok{color:#86efac;font-weight:700}.wide{grid-column:span 3}img{width:100%;border-radius:10px;background:white}@media(max-width:720px){.grid{grid-template-columns:1fr}.wide{grid-column:span 1}} | |
| </style></head><body><div class="wrap"><h1>Distributed Direct Preference Optimization</h1><div class="sub">FedDPO, staleness, lower bounds, DecDPO, and real SHP ablations</div><div class="grid"> | |
| <div class="card"><div class="n">20 / 20</div><div class="ok">release gates pass</div></div><div class="card"><div class="n">600</div><div>real SHP preference pairs</div></div><div class="card"><div class="n">5 × 120</div><div>paper-exact client split</div></div> | |
| <div class="card"><div class="n">72</div><div>FedDPO runs</div></div><div class="card"><div class="n">32</div><div>DecDPO runs</div></div><div class="card"><div class="n">575.65×</div><div>no-mixing consensus penalty</div></div> | |
| <div class="card wide"><img src="outputs/distributed_dpo_audit.png" alt="Distributed DPO theorem and experiment audit"></div></div></div></body></html> | |