Upload folder using huggingface_hub

README.md

@@ -1,10 +1,19 @@
 ---
-title: Ncm Chat Experimental
-emoji: 🌖
-colorFrom: blue
-colorTo: purple
+title: NCM chat (experimental, rough)
+emoji: 💬
+colorFrom: gray
+colorTo: blue
 sdk: static
 pinned: false
+license: mit
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# NCM chat (experimental, rough)
+
+A small chatbot built on **Next-Concept Modeling (NCM)**, Stanford Hazy Research (arXiv 2610.14162, code github.com/HazyResearch/ncm).
+
+- Base: the released XSum NCM checkpoint (frozen multiscale VQ tokenizer plus prefix-conditioned generator). No chat checkpoint is released, so only the generator was fine-tuned for a short time on short public dialogue and instruction data (everyday-conversations, dolly, alpaca).
+- Runs fully in your browser with onnxruntime-web. No server and no data leaves your device.
+- Weights are stored fp16 and cast to fp32 at load (about 375 MB download, roughly 1 GB RAM). Desktop browser recommended. Phones are untested.
+- Export check: on 8 test prompts, greedy replies from the ONNX model are identical to the PyTorch model, and teacher-forced logits match (max diff 0.0026, 0/1600 argmax mismatches).
+- It is a rough research demo. Replies are short and often wrong or repetitive. It is not a capable assistant.

index.html

@@ -1,19 +1,53 @@
-<!doctype html>
-<html>
-	<head>
-		<meta charset="utf-8" />
-		<meta name="viewport" content="width=device-width" />
-		<title>My static Space</title>
-		<link rel="stylesheet" href="style.css" />
-	</head>
-	<body>
-		<div class="card">
-			<h1>Welcome to your static Space!</h1>
-			<p>You can modify this app directly by editing <i>index.html</i> in the Files and versions tab.</p>
-			<p>
-				Also don't forget to check the
-				<a href="https://huggingface.co/docs/hub/spaces" target="_blank">Spaces documentation</a>.
-			</p>
-		</div>
-	</body>
-</html>
+<!doctype html><html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>NCM chat (experimental, rough)</title>
+<style>
+body{font-family:system-ui,sans-serif;margin:0;background:#111;color:#eee;display:flex;flex-direction:column;height:100dvh}
+header{padding:10px 14px;background:#1b1b1b;font-size:13px;line-height:1.35}
+header b{font-size:15px}
+#log{flex:1;overflow:auto;padding:12px;display:flex;flex-direction:column;gap:8px}
+.m{max-width:85%;padding:8px 12px;border-radius:14px;white-space:pre-wrap;word-wrap:break-word}
+.u{align-self:flex-end;background:#2b5fd9}.b{align-self:flex-start;background:#2a2a2a}.s{align-self:center;color:#aaa;font-size:12px}
+form{display:flex;gap:8px;padding:10px;background:#1b1b1b}
+input{flex:1;font-size:16px;padding:10px;border-radius:10px;border:1px solid #444;background:#222;color:#eee}
+button{font-size:16px;padding:10px 16px;border-radius:10px;border:0;background:#2b5fd9;color:#fff}button:disabled{opacity:.5}
+</style></head><body>
+<header><b>NCM chat</b> (experimental, rough) &middot; Next-Concept Modeling, Stanford Hazy Research (arXiv 2610.14162). Runs fully in your browser: the released XSum NCM generator, briefly fine-tuned on short dialogue, fp16-stored-weights ONNX, fp32 compute, verified identical to the PyTorch model on test prompts. Replies are short and often wrong or garbled. Not a capable assistant. First load downloads the model (<span id="mb">~375</span> MB, needs roughly 1 GB free RAM; desktop browser recommended).</header>
+<div id="log"></div>
+<form id="f"><input id="t" autocomplete="off" placeholder="Loading model..." disabled><button id="b" disabled>Send</button></form>
+<script type="module">
+import { AutoTokenizer } from "https://cdn.jsdelivr.net/npm/@huggingface/transformers@3.0.2";
+import * as ort from "https://cdn.jsdelivr.net/npm/onnxruntime-web@1.20.1/dist/ort.min.mjs";
+const MODEL="ncm_fp16w.onnx", P=128, EOS=50256, PAD=50257, PPAD=50257, TOTAL=200, V=4096, T=0.7;
+const log=document.getElementById("log"), inp=document.getElementById("t"), btn=document.getElementById("b");
+const add=(c,txt)=>{const d=document.createElement("div");d.className="m "+c;d.textContent=txt;log.appendChild(d);log.scrollTop=1e9;return d};
+let tok,sess,hist=[];
+async function load(){
+  const s=add("s","Loading tokenizer and model...");
+  tok=await AutoTokenizer.from_pretrained("Xenova/gpt2");
+  ort.env.wasm.wasmPaths="https://cdn.jsdelivr.net/npm/onnxruntime-web@1.20.1/dist/";
+  ort.env.wasm.numThreads=crossOriginIsolated?Math.min(4,navigator.hardwareConcurrency||2):1;
+  const buf=await (await fetch(MODEL)).arrayBuffer();
+  document.getElementById("mb").textContent=Math.round(buf.byteLength/1e6);
+  sess=await ort.InferenceSession.create(buf,{executionProviders:["wasm"]});
+  s.textContent="Model ready. Say something."; inp.disabled=btn.disabled=false; inp.placeholder="Type a message"; inp.focus();
+}
+function gumbel(n){const a=new Float32Array(n);for(let i=0;i<n;i++){const u=Math.random()*(1-1e-7)+1e-7;a[i]=-Math.log(-Math.log(u));}return a}
+async function reply(){
+  const turns=hist.slice(-4); let p=""; for(const [r,t] of turns) p+=(r==="u"?"User: ":"Assistant: ")+t+"\n"; p+="Assistant:";
+  let ids=Array.from(tok.encode(p)).slice(-P); while(ids.length<P) ids.push(PPAD);
+  const prefix=new ort.Tensor("int64",BigInt64Array.from(ids.map(BigInt)),[1,P]);
+  const noise=new ort.Tensor("float32",gumbel(TOTAL*V),[1,TOTAL,V]);
+  const temp=new ort.Tensor("float32",new Float32Array([T]),[]);
+  const out=await sess.run({prefix,noise,temp}); let o=Array.from(out.ids.data,Number);
+  const e=o.indexOf(EOS); if(e>=0) o=o.slice(0,e); o=o.filter(x=>x!==PAD&&x!==EOS);
+  return tok.decode(o,{skip_special_tokens:true}).trim()||"(empty reply)";
+}
+document.getElementById("f").addEventListener("submit",async ev=>{
+  ev.preventDefault(); const m=inp.value.trim(); if(!m) return; inp.value=""; add("u",m); hist.push(["u",m]);
+  inp.disabled=btn.disabled=true; const w=add("b","..."); const t0=performance.now();
+  try{const r=await reply(); w.textContent=r; hist.push(["a",r]); w.title=((performance.now()-t0)/1000).toFixed(1)+"s";}
+  catch(e){w.textContent="Error: "+e.message}
+  inp.disabled=btn.disabled=false; inp.focus();
+});
+load().catch(e=>add("s","Load failed: "+e.message));
+</script></body></html>

ncm_fp16w.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:346557046971e33333aae054cb4d1e07d047778abf3340ddb347038b114988cb
+size 374851158