AI
模型更新

Perplexity 發布 pplx-embed-v2-context-9b-preview：可檢索答案及其佐證的語境嵌入模型

2026年10月1日 03:23
站內 AI 整理稿

Perplexity Research and turbopuffer have released pplx-embed-v2-context-9b-preview, a contextual embedding model for RAG pipelines.Each chunk is embedded with the full document in view.The real change is the training signal.

The model learns to retrieve the answer along with the context needed to verify it, not one ‘gold passage.’ P Is it deployable?Yes, as a self-hosted preview.Weights are on Hugging Face under the MIT license.Loading requires transformers>=5.4.0 with trustremotecode=True.

It is not yet on the Perplexity API.The model card notes that weights and interface may change without backward compatibility.Why the gold passage falls short RAG systems split long documents into chunks.A chunk often depends on an entity, heading, or definition stated elsewhere.

Contextual models address this with late chunking.The document is encoded in one pass, then pooled per chunk.Training, however, usually marks one gold chunk per query.Every other chunk becomes a negative, including the sentences that make the answer checkable.Perplexity lists 3 more problems.

Binary labels give a coarse signal.LLM annotation cost grows linearly with dataset size.Labels are also tied to one chunking strategy.How the training works The teacher is Perplexity’s query-aware context compression model.It reads the query and document together and scores every token.

Chunk relevance: the mean of the top n token scores inside each chunk.Soft target: a temperature-scaled softmax over chunks in the positive document.Chunks in other documents get zero.Distillation loss: forward KL divergence between teacher and student distributions.

Document loss: InfoNCE, where a document scores as its best chunk, inspired by ColBERT’s MaxSim.Each batch samples a random chunking strategy.Chunks are separated by a learned <|chunk_sep|> token and mean-pooled.The teacher runs only during training, so inference adds no latency or storage.

The model starts from an in-house 9B ColBERT retrieval model.A linear projection outputs 2048 dimensions.Matryoshka training also supports 1024 dimensions.Quantization-aware training enables native int8 embeddings.The release is a model soup of several checkpoints.

Training used roughly 430 datasets covering over 50 languages, with no ConTEB data.

Interactive explainer pplx-embed-v2 Contextual Embedding Explainer #pcx{ --ink:#091717; --ink2:#0f2324; --panel:#13343B; --line:#2E565E; --teal:#20808D; --cyan:#1FB8CD; --paper:#FBFAF4; --muted:#9fb7b8; --warn:#e8a33d; --bad:#c9655a; --mtp:#76B900; } #pcx,#pcx {box-sizing:border-box} #pcx{color:var(--paper);font-family:Inter,ui-sans-serif,system-ui,-apple-system,"Segoe UI",Roboto,Arial,sans-serif;line-height:normal;text-align:left} #pcx{max-width:860px;margin:0 auto;padding:22px 18px 18px;background:var(--ink)} #pcx .

hdr{display:flex;align-items:center;gap:12px;margin-bottom:6px} #pcx .

logo{width:34px;height:34px;border-radius:9px;background:linear-gradient(135deg,var(--teal),var(--cyan));display:grid;place-items:center;font-weight:800;color:var(--ink);font-size:15px;flex:none} #pcx h1{font-size:19px;line-height:1.25;margin:0;font-weight:700} #pcx .

sub{color:var(--muted);font-size:13px;margin:4px 0 16px;line-height:1.5} #pcx .tabs{display:flex;gap:6px;flex-wrap:wrap;margin-bottom:14px} #pcx .

tab{background:var(--ink2);border:1px solid var(--line);color:var(--muted);padding:8px 12px;border-radius:999px;font-size:13px;cursor:pointer;transition:all .25s} #pcx .tab b{color:var(--cyan);margin-right:6px;font-weight:700} #pcx .tab.

on{background:var(--teal);border-color:var(--teal);color:var(--paper)} #pcx .tab.on b{color:var(--paper)} #pcx .pane{display:none;animation:pcxfade .45s ease} #pcx .pane.on{display:block} @keyframes pcxfade{from{opacity:0;transform:translateY(6px)}to{opacity:1;transform:none}} #pcx .

card{background:var(--ink2);border:1px solid var(--line);border-radius:14px;padding:14px} #pcx .lead{font-size:14px;line-height:1.6;margin:0 0 12px;color:#dfe9e8} #pcx .

q{display:flex;gap:8px;align-items:center;background:var(--panel);border:1px solid var(--line);border-radius:10px;padding:10px 12px;font-size:13.5px;margin-bottom:12px} #pcx .q .

pill{background:var(--cyan);color:var(--ink);font-weight:700;font-size:11px;padding:3px 8px;border-radius:999px;flex:none} #pcx .ctrl{display:flex;gap:8px;flex-wrap:wrap;align-items:center;margin-bottom:12px} #pcx .

seg{display:inline-flex;border:1px solid var(--line);border-radius:10px;overflow:hidden} #pcx .seg button{background:transparent;border:0;color:var(--muted);padding:8px 12px;font-size:13px;cursor:pointer} #pcx .seg button.on{background:var(--teal);color:var(--paper)} #pcx .

btn{background:var(--cyan);color:var(--ink);border:0;border-radius:10px;padding:9px 14px;font-weight:700;font-size:13px;cursor:pointer;transition:transform .15s} #pcx .btn:active{transform:scale(.97)} #pcx .docs{display:grid;grid-template-columns:repeat(3,1fr);gap:10px} #pcx .

doc{background:var(--ink);border:1px solid var(--line);border-radius:12px;padding:10px;position:relative;transition:border-color .4s, box-shadow .4s} #pcx .doc.win{border-color:var(--cyan);box-shadow:0 0 0 1px var(--cyan),0 0 22px rgba(31,184,205,.25)} #pcx .

doc h4{margin:0 0 8px;font-size:12px;color:var(--muted);font-weight:600;letter-spacing:.02em} #pcx .s{font-size:12px;line-height:1.45;padding:5px 7px;border-radius:7px;margin-bottom:4px;border:1px solid transparent;transition:all .45s;position:relative} #pcx .s.dim{opacity:.38} #pcx .s.

ans{background:rgba(31,184,205,.16);border-color:var(--cyan)} #pcx .s.ev{background:rgba(32,128,141,.22);border-color:var(--teal)} #pcx .s.amb{background:rgba(232,163,61,.14);border-color:var(--warn)} #pcx .s .tag{position:absolute;right:6px;top:-8px;font-size:9.

5px;font-weight:700;padding:1px 6px;border-radius:999px;background:var(--cyan);color:var(--ink);opacity:0;transform:scale(.6);transition:all .35s} #pcx .s.ans .tag,#pcx .s.ev .tag,#pcx .s.amb .tag{opacity:1;transform:none} #pcx .s.ev .tag{background:var(--teal);color:var(--paper)} #pcx .s.amb .

tag{background:var(--warn);color:var(--ink)} #pcx .score{margin-top:8px;height:8px;background:#0b1e1f;border-radius:99px;overflow:hidden} #pcx .score i{display:block;height:100%;width:0;background:var(--cyan);border-radius:99px;transition:width .9s cubic-bezier(.2,.8,.2,1)} #pcx .

sv{font-size:11px;color:var(--muted);margin-top:4px;font-variant-numeric:tabular-nums} #pcx .flow{position:absolute;left:0;top:0;width:100%;height:100%;pointer-events:none;overflow:visible} #pcx .flow path{fill:none;stroke:var(--cyan);stroke-width:1.6;stroke-dasharray:5 5;opacity:0} #pcx .flow.

go path{opacity:.85;animation:pcxdash 1.2s linear infinite} @keyframes pcxdash{to{stroke-dashoffset:-20}} #pcx .verdict{margin-top:12px;font-size:13.5px;line-height:1.55;padding:10px 12px;border-radius:10px;background:var(--panel);border-left:3px solid var(--cyan);min-height:44px} #pcx .

note{font-size:11.5px;color:var(--muted);margin-top:10px;line-height:1.5} #pcx / tab 2 */ .toks{display:flex;flex-wrap:wrap;gap:4px;margin:6px 0 4px} #pcx .tk{font-size:12px;padding:4px 6px;border-radius:6px;background:#0b1e1f;color:#cfe0df;transition:background .6s, color .6s} #pcx .

sep{width:2px;align-self:stretch;background:var(--warn);border-radius:2px;margin:0 3px;transition:opacity .3s} #pcx .row2{display:grid;grid-template-columns:1fr 1fr;gap:10px;margin-top:12px} #pcx .bars .

b{display:grid;grid-template-columns:70px 1fr 46px;gap:8px;align-items:center;font-size:12px;margin-bottom:6px} #pcx .bars .t{height:12px;background:#0b1e1f;border-radius:6px;overflow:hidden} #pcx .bars .t i{display:block;height:100%;width:0;border-radius:6px;transition:width .8s cubic-bezier(.2,.

8,.2,1)} #pcx .bars .num{text-align:right;color:var(--muted);font-variant-numeric:tabular-nums} #pcx .ttl{font-size:12.5px;font-weight:700;margin:0 0 8px;color:var(--paper)} #pcx .ttl span{color:var(--muted);font-weight:500} #pcx input[type=range]{accent-color:var(--cyan);width:160px} #pcx .

lbl{font-size:

原始來源：MarkTechPost AI ↗
查看原始來源

Related

相關文章

IT之家模型更新

OpenAI 稱其遭遇有組織蒸餾，將矛頭指向月之暗面

作者：潞源 責編：潞源 評論： 感謝網友 愚公騎馬、咩咩洋 的線索投遞！10 月 1 日消息，當地時間 9 月 30 日，OpenAI 在官網發文，稱其近期遭遇一場有組織的蒸餾活動。OpenAI 在文中表示，該活動最初始於 7 月第一週，符合對抗式蒸餾（adversarial distillation）之特徵，即系統性未經授權地利用某模型輸出或推理過程，幫助訓練、復現或改進另一模型。

23 小時前
Hugging Face Blog模型更新

推出 Olmo-core 3：為大型 MoE 打造的開放、可擴展訓練基礎架構

今日我們正式釋出 Olmo-core 3，這是我們大型語言模型開發框架的重大升級，核心亮在於重新設計的開放式混合專家（MoE）訓練系統。Olmo-core 3 旨在將 MoE 訓練規模擴展至兆級參數，同時維持運算效率。該系統是下一代 Olmo 的核心基礎之一，亦體現我們持續開放每款新模型背後工具與訓練基礎架構的承諾。

1 天前
MarkTechPost AI模型更新

NVIDIA 發布 Kumo Tabular：開放表格基礎模型，單次前向傳播即可預測新資料列

NVIDIA 推出 Kumo Tabular，這是一系列用於分類與回歸任務的新型表格基礎模型（TFM）。若您曾接觸過 TabPFN 或 TabICL，對此架構應不陌生。該模型以已標記的資料列作為上下文，並在單次前向傳播中預測新資料列，無需訓練、無需超參數調整，也無需特徵工程。Kumo Tabular 提供 Small、Medium 與 Large 三種版本，參數量約從 2,800 萬至 2.15 億不等，並透過 NVIDIA 開源的 structured-data-models（SDM）函式庫運行。此模型可直接部署，其權重採用 OpenMDW-1.1 授權，允許商業使用；SDM 程式碼則以 Apache-2.0 授權發布，需搭配 Python 3.11 以上版本與 PyTorch 2.7 以上版本，範例程式以 CUDA GPU 為目標環境。SDM 函式庫的附加價值在於，它是一個專為結構化資料設計的 GPU 原生函式庫。

1 天前