Clean the model mirror

Removes ~11.6GB of weights that should not be published:
- sixview_v1/v2 checkpoints. sixview_v2 is the pre-standardisation head that
scored WORSE than applying no adapter (and worse than random noise); leaving
it downloadable invites someone to use it. Their metrics.jsonl stay for
provenance.
- duplicate all_lora/head.best.pt (same bytes as all_lora_best_cpt.pth) and a
stray h200_best/ copy.
- 49 files of snapshot_download bookkeeping under data/**/.cache/.

Keeps the two checkpoints worth having: all_lora_best_cpt.pth (merged corpus,
best) and h200_run/head.best.pt (prose-only, fallback).

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

.gitignore

@@ -63,3 +63,6 @@ app/next-env.d.ts
 # Local scratch training run (checkpoint rewritten every step — never commit)
 runs/local_stdfix/
 runs/stdfix_test/
+
+# HF-mirror hygiene: snapshot_download bookkeeping, never useful to publish
+data/**/.cache/

app/engine/generate_and_merge.py

@@ -45,8 +45,34 @@ _DTYPES = {"bfloat16": torch.bfloat16, "float16": torch.float16,
            "float32": torch.float32}
 
 
+def _assert_real_checkpoint(path: Path) -> None:
+    """Fail with an actionable message when the file is a git-LFS pointer.
+
+    A pointer is ~130 bytes of text starting "version https://git-lfs...".
+    Handed to torch.load it surfaces as `Unsupported operand 118` or
+    `invalid load key, 'v'` -- opcode 'v', the first byte of "version" -- which
+    reads like a corrupt checkpoint rather than a file that was never fetched.
+    Any `GIT_LFS_SKIP_SMUDGE=1` clone or pull leaves checkpoints in this state.
+    """
+    if not path.exists():
+        raise FileNotFoundError(f"checkpoint not found: {path}")
+    if path.stat().st_size < 5000:
+        head = path.read_bytes()[:64]
+        if head.startswith(b"version https://git-lfs"):
+            raise RuntimeError(
+                f"{path} is a git-LFS pointer, not the checkpoint "
+                f"({path.stat().st_size} bytes). Fetch it with:\n"
+                f"    git lfs pull --include=\"{path}\""
+            )
+
+
 def load_head(checkpoint: Path):
-    ckpt = torch.load(checkpoint, map_location="cpu")
+    _assert_real_checkpoint(Path(checkpoint))
+    # weights_only=False: these checkpoints carry the run's config/args dicts,
+    # not just tensors, and torch>=2.6 defaults the strict unpickler on --
+    # which rejects them ("Unsupported operand"). They are produced by this
+    # project's own training script, so loading them fully is intended.
+    ckpt = torch.load(checkpoint, map_location="cpu", weights_only=False)
     cfg = ckpt["config"]
     head = MemoryLoRAHead(
         input_dim=cfg["input_dim"],

app/engine/serve_fallback.py

@@ -155,6 +155,16 @@ class Engine:
     def stream(self, prompt: str, max_new_tokens: int, temperature: float,
                top_p: float, stop: list[str] | None = None,
                use_adapter: bool = True):
+        """Stream tokens, with the adapter toggled for the WHOLE operation.
+
+        The enable/disable must wrap the entire generator -- both the worker
+        thread and the consumption of the streamer -- and the thread must be
+        joined before the context exits. PEFT's disable_adapter() flips state on
+        the shared model, so if the context closed while the next request was
+        already starting, that request would silently run with the wrong
+        adapter state. Putting the context inside the worker thread did exactly
+        that: base and adapted both came back as base.
+        """
         from transformers import TextIteratorStreamer
         enc = self.tok(prompt, return_tensors="pt").to(self.device)
         streamer = TextIteratorStreamer(
@@ -167,29 +177,35 @@ class Engine:
             pad_token_id=self.tok.pad_token_id or self.tok.eos_token_id,
             streamer=streamer,
         )
-        def _run():
-            with self._maybe_off(use_adapter):
-                self.model.generate(**kwargs)
-        threading.Thread(target=_run, daemon=True).start()
-
         stops = self._stops(stop)
-        # A stop sequence can straddle two streamed pieces, so emit only the
-        # part of the buffer that can no longer become part of a stop string.
         hold = max(len(s) for s in stops)
-        buf, emitted = "", 0
-        for piece in streamer:
-            buf += piece
-            cut, hit = self._truncate_at_stop(buf, stops)
-            if hit:
-                if len(cut) > emitted:
-                    yield cut[emitted:]
-                return
-            safe = max(0, len(buf) - hold)
-            if safe > emitted:
-                yield buf[emitted:safe]
-                emitted = safe
-        if len(buf) > emitted:
-            yield buf[emitted:]
+
+        with self._maybe_off(use_adapter):
+            worker = threading.Thread(
+                target=self.model.generate, kwargs=kwargs, daemon=True)
+            worker.start()
+            try:
+                # A stop sequence can straddle two streamed pieces, so emit only
+                # the part of the buffer that can no longer become part of one.
+                buf, emitted, hit = "", 0, False
+                for piece in streamer:
+                    buf += piece
+                    cut, hit = self._truncate_at_stop(buf, stops)
+                    if hit:
+                        if len(cut) > emitted:
+                            yield cut[emitted:]
+                        break
+                    safe = max(0, len(buf) - hold)
+                    if safe > emitted:
+                        yield buf[emitted:safe]
+                        emitted = safe
+                if not hit and len(buf) > emitted:
+                    yield buf[emitted:]
+            finally:
+                # Drain and join so the model is idle before adapter state flips.
+                for _ in streamer:
+                    pass
+                worker.join(timeout=120)
 
 
 ENGINE: Engine | None = None

app/src/app/api/jobs/[jobId]/route.ts

@@ -0,0 +1,41 @@
+import { NextRequest, NextResponse } from "next/server";
+import { readStatus } from "@/lib/jobs";
+import { rmSync, existsSync } from "node:fs";
+import path from "node:path";
+
+export const dynamic = "force-dynamic";
+
+const WORKSPACES_DIR = path.join(process.cwd(), ".workspaces");
+
+/** Delete a build: stop its inference server, then remove its workspace. */
+export async function DELETE(
+  _req: NextRequest,
+  { params }: { params: Promise<{ jobId: string }> }
+) {
+  const { jobId } = await params;
+  // Reject anything that could escape the workspaces directory.
+  if (!/^[a-f0-9]{6,32}$/i.test(jobId)) {
+    return NextResponse.json({ error: "invalid job id" }, { status: 400 });
+  }
+  const dir = path.join(WORKSPACES_DIR, jobId);
+  if (!existsSync(dir)) {
+    return NextResponse.json({ error: "not found" }, { status: 404 });
+  }
+
+  // Free the ~10GB the server holds before removing the adapter it loaded.
+  const status = readStatus(jobId);
+  const pid = status?.server?.pid;
+  if (pid) {
+    try {
+      process.kill(pid, "SIGTERM");
+    } catch {
+      /* already gone */
+    }
+  }
+  try {
+    rmSync(dir, { recursive: true, force: true });
+  } catch (e) {
+    return NextResponse.json({ error: String(e) }, { status: 500 });
+  }
+  return NextResponse.json({ ok: true, jobId });
+}

app/src/app/api/jobs/[jobId]/server/route.ts

@@ -0,0 +1,118 @@
+import { NextRequest, NextResponse } from "next/server";
+import { readStatus } from "@/lib/jobs";
+import { spawn } from "node:child_process";
+import { existsSync, readFileSync, writeFileSync } from "node:fs";
+import path from "node:path";
+
+export const dynamic = "force-dynamic";
+
+const APP_DIR = path.resolve(process.cwd());
+const ENGINE_DIR = path.join(APP_DIR, "engine");
+const WORKSPACES_DIR = path.join(APP_DIR, ".workspaces");
+const PYTHON = process.env.MLORA_PYTHON || "python3";
+
+function statusPath(jobId: string) {
+  return path.join(WORKSPACES_DIR, jobId, "status.json");
+}
+
+/**
+ * Stop or restart a job's inference server.
+ *
+ * A transformers server that has been generating for a while slows down badly
+ * (MPS allocator churn, accumulated KV state), and the only reliable cure is a
+ * fresh process. Restarting reloads the same base+adapter, so the endpoint
+ * comes back on the same port with identical behaviour — no rebuild needed.
+ */
+export async function POST(
+  req: NextRequest,
+  { params }: { params: Promise<{ jobId: string }> }
+) {
+  const { jobId } = await params;
+  if (!/^[a-f0-9]{6,32}$/i.test(jobId)) {
+    return NextResponse.json({ error: "invalid job id" }, { status: 400 });
+  }
+  const dir = path.join(WORKSPACES_DIR, jobId);
+  if (!existsSync(dir)) {
+    return NextResponse.json({ error: "not found" }, { status: 404 });
+  }
+
+  let action = "restart";
+  try {
+    action = (await req.json())?.action || "restart";
+  } catch {
+    /* default */
+  }
+
+  const status = readStatus(jobId);
+  const oldPid = status?.server?.pid;
+
+  // Kill first in both cases — a restart must not leave two processes fighting
+  // over the same port.
+  if (oldPid) {
+    try {
+      process.kill(oldPid, "SIGTERM");
+    } catch {
+      /* already gone */
+    }
+    await new Promise((r) => setTimeout(r, 1200));
+    try {
+      process.kill(oldPid, 0);
+      process.kill(oldPid, "SIGKILL"); // still alive: force it
+    } catch {
+      /* exited cleanly */
+    }
+  }
+
+  const raw = existsSync(statusPath(jobId))
+    ? JSON.parse(readFileSync(statusPath(jobId), "utf8"))
+    : {};
+
+  if (action === "stop") {
+    raw.state = "error";
+    raw.error = "server stopped — press Restart to bring it back";
+    raw.updated_at = Date.now() / 1000;
+    writeFileSync(statusPath(jobId), JSON.stringify(raw, null, 2));
+    return NextResponse.json({ ok: true, action: "stop" });
+  }
+
+  const port = status?.server?.port || 8000;
+  const adapter = path.join(dir, "adapter");
+  const merged = path.join(dir, "merged");
+  const args = [
+    path.join(ENGINE_DIR, "serve_fallback.py"),
+    "--port", String(port),
+    "--served-name", "memory-lora",
+    "--job", jobId,
+  ];
+  if (existsSync(adapter)) {
+    args.push("--adapter", adapter);
+  } else if (existsSync(merged)) {
+    args.push("--model", merged);
+  } else {
+    return NextResponse.json(
+      { error: "no adapter or merged model in this workspace" },
+      { status: 409 }
+    );
+  }
+  if (process.env.MLORA_DEVICE) args.push("--device", process.env.MLORA_DEVICE);
+
+  const child = spawn(PYTHON, args, {
+    cwd: ENGINE_DIR,
+    detached: true,
+    stdio: "ignore",
+    env: { ...process.env, MLORA_WORKSPACES: WORKSPACES_DIR },
+  });
+  child.unref();
+
+  // serve_fallback --job rewrites status.json with its own pid once the model
+  // is loaded; mark it as loading meanwhile so the UI doesn't claim it's ready.
+  raw.state = "running";
+  raw.stage = "serve";
+  raw.error = null;
+  raw.server = { backend: "fallback", pid: child.pid, port, mode: "lora" };
+  raw.pipeline_pid = child.pid;
+  raw.updated_at = Date.now() / 1000;
+  writeFileSync(statusPath(jobId), JSON.stringify(raw, null, 2));
+
+  return NextResponse.json({ ok: true, action: "restart", pid: child.pid, port });
+}

app/src/app/compare/page.tsx

@@ -1,270 +1,5 @@
 "use client";
 
-import { useEffect, useRef, useState } from "react";
-
-type Job = {
-  job_id: string;
-  repo_url: string;
-  stages: string[];
-  stage: string;
-  state: "running" | "ready" | "error";
-  error?: string | null;
-  server?: { backend: string; port: number; mode?: string };
-};
-
-type Turn = { role: "user" | "assistant"; base?: string; adapted?: string };
-
-const STAGE_LABEL: Record<string, string> = {
-  clone: "clone",
-  embed: "embed",
-  model: "base model",
-  generate: "adapter",
-  merge: "merge",
-  serve: "serve",
-  ready: "ready",
-};
-
-// Questions that expose repo knowledge rather than general fluency — the base
-// model can bluff a plausible answer to all of them, which is the point.
-const SUGGESTED = [
-  "What is the core purpose of this repository?",
-  "What testing framework does this repository use?",
-  "How is this project built and packaged?",
-  "What are the main modules and how do they fit together?",
-];
-
-export default function Compare() {
-  const [url, setUrl] = useState("");
-  const [job, setJob] = useState<Job | null>(null);
-  const [jobs, setJobs] = useState<Job[]>([]);
-  const [turns, setTurns] = useState<Turn[]>([]);
-  const [q, setQ] = useState("");
-  const [busy, setBusy] = useState(false);
-  const scroller = useRef<HTMLDivElement>(null);
-
-  useEffect(() => {
-    const poll = async () => {
-      try {
-        const r = await fetch("/api/build");
-        const d = await r.json();
-        setJobs(d.jobs || []);
-        setJob((cur) =>
-          cur ? (d.jobs || []).find((j: Job) => j.job_id === cur.job_id) || cur : cur
-        );
-      } catch {
-        /* ignore */
-      }
-    };
-    poll();
-    const t = setInterval(poll, 2000);
-    return () => clearInterval(t);
-  }, []);
-
-  useEffect(() => {
-    scroller.current?.scrollTo({ top: scroller.current.scrollHeight, behavior: "smooth" });
-  }, [turns]);
-
-  const build = async () => {
-    if (!url.trim()) return;
-    setBusy(true);
-    try {
-      const r = await fetch("/api/build", {
-        method: "POST",
-        headers: { "content-type": "application/json" },
-        body: JSON.stringify({ repoUrl: url.trim() }),
-      });
-      const d = await r.json();
-      setJob({
-        job_id: d.jobId,
-        repo_url: url.trim(),
-        stages: ["clone", "embed", "model", "generate", "serve", "ready"],
-        stage: "clone",
-        state: "running",
-      });
-      setTurns([]);
-      setUrl("");
-    } finally {
-      setBusy(false);
-    }
-  };
-
-  // Both sides hit the same server; only the model id differs, so the
-  // comparison is the adapter and nothing else.
-  const askOne = async (model: string, question: string): Promise<string> => {
-    const r = await fetch("/v1/chat/completions", {
-      method: "POST",
-      headers: { "content-type": "application/json", "x-mlora-job": job!.job_id },
-      body: JSON.stringify({
-        model,
-        max_tokens: 160,
-        messages: [{ role: "user", content: question }],
-      }),
-    });
-    if (!r.ok) return `[error ${r.status}]`;
-    const d = await r.json();
-    return d.choices?.[0]?.message?.content?.trim() || "[empty]";
-  };
-
-  const ask = async (question?: string) => {
-    const text = (question ?? q).trim();
-    if (!text || !job || job.state !== "ready" || busy) return;
-    setBusy(true);
-    setQ("");
-    setTurns((t) => [...t, { role: "user", base: text }, { role: "assistant" }]);
-    try {
-      const [base, adapted] = await Promise.all([
-        askOne("base", text),
-        askOne(`memory-lora:${job.job_id}`, text),
-      ]);
-      setTurns((t) => {
-        const c = [...t];
-        c[c.length - 1] = { role: "assistant", base, adapted };
-        return c;
-      });
-    } finally {
-      setBusy(false);
-    }
-  };
-
-  const ready = job?.state === "ready";
-  const idx = job ? job.stages?.indexOf(job.stage) ?? 0 : 0;
-
-  return (
-    <div className="cmp">
-      <header className="cmp-head">
-        <div>
-          <div className="eyebrow">Memory-LoRA · live comparison</div>
-          <h1>Same model. Same prompt. One has read your repo.</h1>
-        </div>
-        {job && (
-          <span className={`badge ${job.state}`}>
-            {job.state === "ready" ? "endpoint live" : job.state}
-          </span>
-        )}
-      </header>
-
-      <div className="bar">
-        <input
-          type="text"
-          placeholder="https://github.com/owner/repo"
-          value={url}
-          onChange={(e) => setUrl(e.target.value)}
-          onKeyDown={(e) => e.key === "Enter" && build()}
-        />
-        <button onClick={build} disabled={busy || !url.trim()}>
-          {busy ? "Starting…" : "Build adapter"}
-        </button>
-      </div>
-
-      {jobs.length > 0 && (
-        <div className="recent">
-          <span className="eyebrow">Ready</span>
-          {jobs
-            .filter((j) => j.state === "ready")
-            .slice(0, 4)
-            .map((j) => (
-              <button
-                key={j.job_id}
-                className={`chip ${job?.job_id === j.job_id ? "on" : ""}`}
-                onClick={() => {
-                  setJob(j);
-                  setTurns([]);
-                }}
-              >
-                {j.repo_url.replace("https://github.com/", "")}
-              </button>
-            ))}
-        </div>
-      )}
-
-      {job && !ready && (
-        <div className="progress">
-          <div className="stages">
-            {(job.stages || []).map((s, i) => (
-              <span
-                key={s}
-                className={`stage ${i < idx ? "done" : ""} ${
-                  i === idx && job.state === "running" ? "active" : ""
-                }`}
-              >
-                {STAGE_LABEL[s] || s}
-              </span>
-            ))}
-          </div>
-          {job.state === "error" ? (
-            <p className="err">✗ {job.error}</p>
-          ) : (
-            <p className="hint">
-              Cloning, embedding six views, and generating a LoRA adapter — about a minute.
-            </p>
-          )}
-        </div>
-      )}
-
-      {ready && (
-        <>
-          <div className="panes">
-            <div className="pane-head base">
-              <span className="dot" /> Frozen Gemma-4-E2B
-              <em>no repo knowledge</em>
-            </div>
-            <div className="pane-head adapted">
-              <span className="dot" /> + generated adapter
-              <em>{job!.repo_url.replace("https://github.com/", "")}</em>
-            </div>
-          </div>
-
-          <div className="scroll" ref={scroller}>
-            {turns.length === 0 && (
-              <div className="empty">
-                <p>Ask something only this repository can answer.</p>
-                <div className="suggest">
-                  {SUGGESTED.map((s) => (
-                    <button key={s} onClick={() => ask(s)} disabled={busy}>
-                      {s}
-                    </button>
-                  ))}
-                </div>
-              </div>
-            )}
-            {turns.map((t, i) =>
-              t.role === "user" ? (
-                <div key={i} className="q">
-                  {t.base}
-                </div>
-              ) : (
-                <div key={i} className="answers">
-                  <div className="ans base">
-                    {t.base ?? <span className="think">thinking…</span>}
-                  </div>
-                  <div className="ans adapted">
-                    {t.adapted ?? <span className="think">thinking…</span>}
-                  </div>
-                </div>
-              )
-            )}
-          </div>
-
-          <div className="ask">
-            <input
-              type="text"
-              placeholder="Ask about this repository…"
-              value={q}
-              onChange={(e) => setQ(e.target.value)}
-              onKeyDown={(e) => e.key === "Enter" && ask()}
-              disabled={busy}
-            />
-            <button onClick={() => ask()} disabled={busy || !q.trim()}>
-              {busy ? "…" : "Ask both"}
-            </button>
-          </div>
-          <p className="foot">
-            Both answers come from one resident model on the same server — the only
-            difference is whether the generated adapter is switched on. No repository
-            text is in either prompt.
-          </p>
-        </>
-      )}
-    </div>
-  );
-}
+// The demo lives at "/" — this keeps "/compare" working too, so either URL
+// lands on the side-by-side comparison instead of a 404.
+export { default } from "../page";

app/src/app/globals.css

@@ -162,3 +162,20 @@ label.small { color: var(--muted); font-size: 13px; }
              border-radius: 10px; color: var(--text); padding: 13px 15px; font-size: 15px; }
 .ask input:focus { outline: none; border-color: var(--accent); }
 .foot { color: var(--muted); font-size: 12.5px; margin: 12px 0 0; text-align: center; }
+.chip { display: inline-flex; align-items: center; gap: 0; padding: 0; overflow: hidden; }
+.chip-main { background: none; border: none; color: inherit; font: inherit;
+             padding: 5px 6px 5px 12px; cursor: pointer; font-family: var(--mono); }
+.chip-main em { font-style: normal; color: var(--muted); }
+.chip-x { background: none; border: none; color: var(--muted); cursor: pointer;
+          font-size: 15px; line-height: 1; padding: 5px 10px 6px 4px; font-weight: 400; }
+.chip-x:hover { color: var(--err); }
+.clear-all { background: none; border: 1px solid var(--border); color: var(--muted);
+             border-radius: 999px; padding: 5px 12px; font-size: 12.5px; cursor: pointer;
+             font-weight: 400; }
+.clear-all:hover { border-color: var(--err); color: var(--err); }
+button.ghost { background: transparent; border: 1px solid var(--border);
+               color: var(--muted); font-weight: 400; }
+button.ghost:hover:not(:disabled) { border-color: var(--accent); color: var(--text); }
+.head-actions { display: flex; align-items: center; gap: 8px; flex-wrap: wrap; }
+button.small { padding: 5px 12px; font-size: 12.5px; border-radius: 999px; }
+button.ghost.danger:hover:not(:disabled) { border-color: var(--err); color: var(--err); }

app/src/app/page.tsx

@@ -1,6 +1,6 @@
 "use client";
 
-import { useEffect, useState } from "react";
+import { useEffect, useRef, useState } from "react";
 
 type Job = {
   job_id: string;
@@ -9,10 +9,11 @@ type Job = {
   stage: string;
   state: "running" | "ready" | "error";
   error?: string | null;
-  server?: { backend: string; port: number };
-  endpoint?: string;
+  server?: { backend: string; port: number; mode?: string };
 };
 
+type Turn = { role: "user" | "assistant"; base?: string; adapted?: string };
+
 const STAGE_LABEL: Record<string, string> = {
   clone: "clone",
   embed: "embed",
@@ -23,80 +24,37 @@ const STAGE_LABEL: Record<string, string> = {
   ready: "ready",
 };
 
-function ConnectSnippets({ job, base }: { job: Job; base: string }) {
-  const [tab, setTab] = useState<"claude" | "openai" | "curl">("claude");
-  const model = `memory-lora:${job.job_id}`;
-  const claude = `# Claude Code — point it at this repo-personalized model
-export ANTHROPIC_BASE_URL="${base}"
-export ANTHROPIC_API_KEY="local-demo"       # any non-empty value
-export ANTHROPIC_MODEL="${model}"
-claude`;
-  const openai = `# vibe / aider / any OpenAI-compatible CLI
-export OPENAI_BASE_URL="${base}/v1"
-export OPENAI_API_KEY="local-demo"
-export OPENAI_MODEL="${model}"
-# e.g. aider --model openai/${model}
-# e.g. vibe --base-url $OPENAI_BASE_URL --model ${model}`;
-  const curl = `curl ${base}/v1/chat/completions \\
-  -H "content-type: application/json" \\
-  -d '{"model":"${model}","messages":[{"role":"user","content":"What does this repo do?"}]}'`;
-  const text = tab === "claude" ? claude : tab === "openai" ? openai : curl;
-  return (
-    <div style={{ marginTop: 14 }}>
-      <div className="tabs">
-        <div className={`tab ${tab === "claude" ? "on" : ""}`} onClick={() => setTab("claude")}>Claude Code</div>
-        <div className={`tab ${tab === "openai" ? "on" : ""}`} onClick={() => setTab("openai")}>OpenAI CLIs</div>
-        <div className={`tab ${tab === "curl" ? "on" : ""}`} onClick={() => setTab("curl")}>curl</div>
-      </div>
-      <pre>{text}</pre>
-      <p className="hint">
-        Endpoint served by {job.server?.backend || "engine"} on port {job.server?.port}. The app
-        translates the Anthropic <code>/v1/messages</code> API to this model, so Claude Code works
-        directly.
-      </p>
-    </div>
-  );
-}
+// Questions that expose repo knowledge rather than general fluency — the base
+// model can bluff a plausible answer to all of them, which is the point.
+// Short answers keep the demo snappy; the server's stop sequences cut the
+// response at the end of the answer anyway, so this rarely truncates.
+const MAX_TOKENS = 72;
 
-function JobCard({ job, base }: { job: Job; base: string }) {
-  const idx = job.stages?.indexOf(job.stage) ?? 0;
-  return (
-    <div className="card">
-      <div className="jobhead">
-        <span className="repo">{job.repo_url}</span>
-        <span className={`badge ${job.state}`}>{job.state}</span>
-      </div>
-      <div className="stages">
-        {(job.stages || []).map((s, i) => (
-          <span
-            key={s}
-            className={`stage ${i < idx || job.state === "ready" ? "done" : ""} ${
-              i === idx && job.state === "running" ? "active" : ""
-            }`}
-          >
-            {STAGE_LABEL[s] || s}
-          </span>
-        ))}
-      </div>
-      {job.state === "error" && <div className="err">✗ {job.error}</div>}
-      {job.state === "ready" && <ConnectSnippets job={job} base={base} />}
-    </div>
-  );
-}
+const SUGGESTED = [
+  "What is the core purpose of this repository?",
+  "What testing framework does this repository use?",
+  "How is this project built and packaged?",
+  "What are the main modules and how do they fit together?",
+];
 
-export default function Home() {
+export default function Compare() {
   const [url, setUrl] = useState("");
+  const [job, setJob] = useState<Job | null>(null);
   const [jobs, setJobs] = useState<Job[]>([]);
+  const [turns, setTurns] = useState<Turn[]>([]);
+  const [q, setQ] = useState("");
   const [busy, setBusy] = useState(false);
-  const [base, setBase] = useState("");
+  const scroller = useRef<HTMLDivElement>(null);
 
   useEffect(() => {
-    setBase(window.location.origin);
     const poll = async () => {
       try {
         const r = await fetch("/api/build");
         const d = await r.json();
         setJobs(d.jobs || []);
+        setJob((cur) =>
+          cur ? (d.jobs || []).find((j: Job) => j.job_id === cur.job_id) || cur : cur
+        );
       } catch {
         /* ignore */
       }
@@ -106,56 +64,322 @@ export default function Home() {
     return () => clearInterval(t);
   }, []);
 
-  const submit = async () => {
+  useEffect(() => {
+    scroller.current?.scrollTo({ top: scroller.current.scrollHeight, behavior: "smooth" });
+  }, [turns]);
+
+  const build = async () => {
     if (!url.trim()) return;
     setBusy(true);
     try {
-      await fetch("/api/build", {
+      const r = await fetch("/api/build", {
         method: "POST",
         headers: { "content-type": "application/json" },
         body: JSON.stringify({ repoUrl: url.trim() }),
       });
+      const d = await r.json();
+      setJob({
+        job_id: d.jobId,
+        repo_url: url.trim(),
+        stages: ["clone", "embed", "model", "generate", "serve", "ready"],
+        stage: "clone",
+        state: "running",
+      });
+      setTurns([]);
       setUrl("");
     } finally {
       setBusy(false);
     }
   };
 
+  // Both sides hit the same server; only the model id differs, so the
+  // comparison is the adapter and nothing else.
+  //
+  // Streamed, and deliberately sequential: one model serves both sides, so the
+  // server holds a lock and concurrent requests would queue anyway. Streaming
+  // is what makes it feel immediate -- tokens appear as they are produced
+  // instead of after the full answer.
+  const askOne = async (
+    model: string,
+    question: string,
+    onDelta: (chunk: string) => void
+  ): Promise<void> => {
+    const r = await fetch("/v1/chat/completions", {
+      method: "POST",
+      headers: { "content-type": "application/json", "x-mlora-job": job!.job_id },
+      body: JSON.stringify({
+        model,
+        max_tokens: MAX_TOKENS,
+        stream: true,
+        messages: [{ role: "user", content: question }],
+      }),
+    });
+    if (!r.ok || !r.body) {
+      onDelta(`[error ${r.status}]`);
+      return;
+    }
+    const reader = r.body.getReader();
+    const dec = new TextDecoder();
+    let buf = "";
+    for (;;) {
+      const { done, value } = await reader.read();
+      if (done) break;
+      buf += dec.decode(value, { stream: true });
+      const lines = buf.split("\n");
+      buf = lines.pop() || "";
+      for (const line of lines) {
+        const t = line.trim();
+        if (!t.startsWith("data:")) continue;
+        const payload = t.slice(5).trim();
+        if (payload === "[DONE]") continue;
+        try {
+          const d = JSON.parse(payload);
+          const piece = d.choices?.[0]?.delta?.content;
+          if (piece) onDelta(piece);
+        } catch {
+          /* partial frame */
+        }
+      }
+    }
+  };
+
+  const ask = async (question?: string) => {
+    const text = (question ?? q).trim();
+    if (!text || !job || job.state !== "ready" || busy) return;
+    setBusy(true);
+    setQ("");
+    setTurns((t) => [
+      ...t,
+      { role: "user", base: text },
+      { role: "assistant", base: "", adapted: "" },
+    ]);
+    const push = (side: "base" | "adapted") => (chunk: string) =>
+      setTurns((t) => {
+        const c = [...t];
+        const last = { ...c[c.length - 1] };
+        last[side] = (last[side] || "") + chunk;
+        c[c.length - 1] = last;
+        return c;
+      });
+    try {
+      await askOne("base", text, push("base"));
+      await askOne(`memory-lora:${job.job_id}`, text, push("adapted"));
+    } finally {
+      setBusy(false);
+    }
+  };
+
+  const clearChat = () => setTurns([]);
+
+  // A long-lived transformers server degrades badly (allocator churn on MPS);
+  // a fresh process reloads the same base+adapter on the same port.
+  const serverAction = async (action: "stop" | "restart") => {
+    if (!job) return;
+    setBusy(true);
+    try {
+      await fetch(`/api/jobs/${job.job_id}/server`, {
+        method: "POST",
+        headers: { "content-type": "application/json" },
+        body: JSON.stringify({ action }),
+      });
+      const r = await fetch("/api/build");
+      setJobs((await r.json()).jobs || []);
+    } finally {
+      setBusy(false);
+    }
+  };
+
+  const deleteBuild = async (jobId: string) => {
+    await fetch(`/api/jobs/${jobId}`, { method: "DELETE" });
+    if (job?.job_id === jobId) {
+      setJob(null);
+      setTurns([]);
+    }
+    const r = await fetch("/api/build");
+    setJobs((await r.json()).jobs || []);
+  };
+
+  const clearAll = async () => {
+    await Promise.all(jobs.map((j) => fetch(`/api/jobs/${j.job_id}`, { method: "DELETE" })));
+    setJob(null);
+    setTurns([]);
+    setJobs([]);
+  };
+
+  const ready = job?.state === "ready";
+  const idx = job ? job.stages?.indexOf(job.stage) ?? 0 : 0;
+
   return (
-    <div className="wrap">
-      <h1>Memory-LoRA</h1>
-      <p className="sub">
-        Paste a git repository. A hypernetwork reads a 6-view embedding of the codebase and{" "}
-        <b>emits a LoRA adapter</b> for Gemma in one forward pass — no fine-tuning. The adapter is
-        merged into the base model and served behind an OpenAI + Anthropic compatible endpoint, so
-        your coding CLI talks to a model that already knows this repo.
-      </p>
-
-      <div className="card">
-        <div className="row">
-          <input
-            type="text"
-            placeholder="https://github.com/owner/repo"
-            value={url}
-            onChange={(e) => setUrl(e.target.value)}
-            onKeyDown={(e) => e.key === "Enter" && submit()}
-          />
-          <button onClick={submit} disabled={busy}>
-            {busy ? "Starting…" : "Build & serve"}
-          </button>
+    <div className="cmp">
+      <header className="cmp-head">
+        <div>
+          <div className="eyebrow">Memory-LoRA · live comparison</div>
+          <h1>Same model. Same prompt. One has read your repo.</h1>
         </div>
-        <p className="hint">
-          First build downloads the base model (~10&nbsp;GB) and takes a few minutes; later builds
-          reuse it.
-        </p>
+        {job && (
+          <div className="head-actions">
+            <span className={`badge ${job.state}`}>
+              {job.state === "ready" ? "endpoint live" : job.state}
+            </span>
+            <button
+              className="ghost small"
+              onClick={() => serverAction("restart")}
+              disabled={busy}
+              title="Kill and reload the model process — fixes a server that has become slow"
+            >
+              Restart model
+            </button>
+            <button
+              className="ghost small danger"
+              onClick={() => serverAction("stop")}
+              disabled={busy || job.state !== "ready"}
+              title="Kill the model process and free its memory"
+            >
+              Kill
+            </button>
+          </div>
+        )}
+      </header>
+
+      <div className="bar">
+        <input
+          type="text"
+          placeholder="https://github.com/owner/repo"
+          value={url}
+          onChange={(e) => setUrl(e.target.value)}
+          onKeyDown={(e) => e.key === "Enter" && build()}
+        />
+        <button onClick={build} disabled={busy || !url.trim()}>
+          {busy ? "Starting…" : "Build adapter"}
+        </button>
       </div>
 
-      {jobs.length === 0 && (
-        <p className="hint">No builds yet. Paste a repo URL above to start one.</p>
+      {jobs.length > 0 && (
+        <div className="recent">
+          <span className="eyebrow">Ready</span>
+          {jobs.slice(0, 5).map((j) => (
+            <span key={j.job_id} className={`chip ${job?.job_id === j.job_id ? "on" : ""}`}>
+              <button
+                className="chip-main"
+                onClick={() => {
+                  setJob(j);
+                  setTurns([]);
+                }}
+              >
+                {j.repo_url.replace("https://github.com/", "")}
+                {j.state !== "ready" && <em> · {j.state}</em>}
+              </button>
+              <button
+                className="chip-x"
+                title="Delete this build"
+                aria-label={`Delete ${j.repo_url}`}
+                onClick={() => deleteBuild(j.job_id)}
+              >
+                ×
+              </button>
+            </span>
+          ))}
+          {jobs.length > 0 && (
+            <button className="clear-all" onClick={clearAll}>
+              Clear all builds
+            </button>
+          )}
+        </div>
+      )}
+
+      {job && !ready && (
+        <div className="progress">
+          <div className="stages">
+            {(job.stages || []).map((s, i) => (
+              <span
+                key={s}
+                className={`stage ${i < idx ? "done" : ""} ${
+                  i === idx && job.state === "running" ? "active" : ""
+                }`}
+              >
+                {STAGE_LABEL[s] || s}
+              </span>
+            ))}
+          </div>
+          {job.state === "error" ? (
+            <p className="err">✗ {job.error}</p>
+          ) : (
+            <p className="hint">
+              Cloning, embedding six views, and generating a LoRA adapter — about a minute.
+            </p>
+          )}
+        </div>
+      )}
+
+      {ready && (
+        <>
+          <div className="panes">
+            <div className="pane-head base">
+              <span className="dot" /> Frozen Gemma-4-E2B
+              <em>no repo knowledge</em>
+            </div>
+            <div className="pane-head adapted">
+              <span className="dot" /> + generated adapter
+              <em>{job!.repo_url.replace("https://github.com/", "")}</em>
+            </div>
+          </div>
+
+          <div className="scroll" ref={scroller}>
+            {turns.length === 0 && (
+              <div className="empty">
+                <p>Ask something only this repository can answer.</p>
+                <div className="suggest">
+                  {SUGGESTED.map((s) => (
+                    <button key={s} onClick={() => ask(s)} disabled={busy}>
+                      {s}
+                    </button>
+                  ))}
+                </div>
+              </div>
+            )}
+            {turns.map((t, i) =>
+              t.role === "user" ? (
+                <div key={i} className="q">
+                  {t.base}
+                </div>
+              ) : (
+                <div key={i} className="answers">
+                  <div className="ans base">
+                    {t.base ? t.base : <span className="think">…</span>}
+                  </div>
+                  <div className="ans adapted">
+                    {t.adapted ? t.adapted : <span className="think">…</span>}
+                  </div>
+                </div>
+              )
+            )}
+          </div>
+
+          <div className="ask">
+            <input
+              type="text"
+              placeholder="Ask about this repository…"
+              value={q}
+              onChange={(e) => setQ(e.target.value)}
+              onKeyDown={(e) => e.key === "Enter" && ask()}
+              disabled={busy}
+            />
+            <button onClick={() => ask()} disabled={busy || !q.trim()}>
+              {busy ? "…" : "Ask both"}
+            </button>
+            {turns.length > 0 && (
+              <button className="ghost" onClick={clearChat} disabled={busy}>
+                Clear chat
+              </button>
+            )}
+          </div>
+          <p className="foot">
+            Both answers come from one resident model on the same server — the only
+            difference is whether the generated adapter is switched on. No repository
+            text is in either prompt.
+          </p>
+        </>
       )}
-      {jobs.map((j) => (
-        <JobCard key={j.job_id} job={j} base={base} />
-      ))}
     </div>
   );
 }

app/src/app/serve/page.tsx

@@ -0,0 +1,161 @@
+"use client";
+
+import { useEffect, useState } from "react";
+
+type Job = {
+  job_id: string;
+  repo_url: string;
+  stages: string[];
+  stage: string;
+  state: "running" | "ready" | "error";
+  error?: string | null;
+  server?: { backend: string; port: number };
+  endpoint?: string;
+};
+
+const STAGE_LABEL: Record<string, string> = {
+  clone: "clone",
+  embed: "embed",
+  model: "base model",
+  generate: "adapter",
+  merge: "merge",
+  serve: "serve",
+  ready: "ready",
+};
+
+function ConnectSnippets({ job, base }: { job: Job; base: string }) {
+  const [tab, setTab] = useState<"claude" | "openai" | "curl">("claude");
+  const model = `memory-lora:${job.job_id}`;
+  const claude = `# Claude Code — point it at this repo-personalized model
+export ANTHROPIC_BASE_URL="${base}"
+export ANTHROPIC_API_KEY="local-demo"       # any non-empty value
+export ANTHROPIC_MODEL="${model}"
+claude`;
+  const openai = `# vibe / aider / any OpenAI-compatible CLI
+export OPENAI_BASE_URL="${base}/v1"
+export OPENAI_API_KEY="local-demo"
+export OPENAI_MODEL="${model}"
+# e.g. aider --model openai/${model}
+# e.g. vibe --base-url $OPENAI_BASE_URL --model ${model}`;
+  const curl = `curl ${base}/v1/chat/completions \\
+  -H "content-type: application/json" \\
+  -d '{"model":"${model}","messages":[{"role":"user","content":"What does this repo do?"}]}'`;
+  const text = tab === "claude" ? claude : tab === "openai" ? openai : curl;
+  return (
+    <div style={{ marginTop: 14 }}>
+      <div className="tabs">
+        <div className={`tab ${tab === "claude" ? "on" : ""}`} onClick={() => setTab("claude")}>Claude Code</div>
+        <div className={`tab ${tab === "openai" ? "on" : ""}`} onClick={() => setTab("openai")}>OpenAI CLIs</div>
+        <div className={`tab ${tab === "curl" ? "on" : ""}`} onClick={() => setTab("curl")}>curl</div>
+      </div>
+      <pre>{text}</pre>
+      <p className="hint">
+        Endpoint served by {job.server?.backend || "engine"} on port {job.server?.port}. The app
+        translates the Anthropic <code>/v1/messages</code> API to this model, so Claude Code works
+        directly.
+      </p>
+    </div>
+  );
+}
+
+function JobCard({ job, base }: { job: Job; base: string }) {
+  const idx = job.stages?.indexOf(job.stage) ?? 0;
+  return (
+    <div className="card">
+      <div className="jobhead">
+        <span className="repo">{job.repo_url}</span>
+        <span className={`badge ${job.state}`}>{job.state}</span>
+      </div>
+      <div className="stages">
+        {(job.stages || []).map((s, i) => (
+          <span
+            key={s}
+            className={`stage ${i < idx || job.state === "ready" ? "done" : ""} ${
+              i === idx && job.state === "running" ? "active" : ""
+            }`}
+          >
+            {STAGE_LABEL[s] || s}
+          </span>
+        ))}
+      </div>
+      {job.state === "error" && <div className="err">✗ {job.error}</div>}
+      {job.state === "ready" && <ConnectSnippets job={job} base={base} />}
+    </div>
+  );
+}
+
+export default function Home() {
+  const [url, setUrl] = useState("");
+  const [jobs, setJobs] = useState<Job[]>([]);
+  const [busy, setBusy] = useState(false);
+  const [base, setBase] = useState("");
+
+  useEffect(() => {
+    setBase(window.location.origin);
+    const poll = async () => {
+      try {
+        const r = await fetch("/api/build");
+        const d = await r.json();
+        setJobs(d.jobs || []);
+      } catch {
+        /* ignore */
+      }
+    };
+    poll();
+    const t = setInterval(poll, 2000);
+    return () => clearInterval(t);
+  }, []);
+
+  const submit = async () => {
+    if (!url.trim()) return;
+    setBusy(true);
+    try {
+      await fetch("/api/build", {
+        method: "POST",
+        headers: { "content-type": "application/json" },
+        body: JSON.stringify({ repoUrl: url.trim() }),
+      });
+      setUrl("");
+    } finally {
+      setBusy(false);
+    }
+  };
+
+  return (
+    <div className="wrap">
+      <h1>Memory-LoRA</h1>
+      <p className="sub">
+        Paste a git repository. A hypernetwork reads a 6-view embedding of the codebase and{" "}
+        <b>emits a LoRA adapter</b> for Gemma in one forward pass — no fine-tuning. The adapter is
+        merged into the base model and served behind an OpenAI + Anthropic compatible endpoint, so
+        your coding CLI talks to a model that already knows this repo.
+      </p>
+
+      <div className="card">
+        <div className="row">
+          <input
+            type="text"
+            placeholder="https://github.com/owner/repo"
+            value={url}
+            onChange={(e) => setUrl(e.target.value)}
+            onKeyDown={(e) => e.key === "Enter" && submit()}
+          />
+          <button onClick={submit} disabled={busy}>
+            {busy ? "Starting…" : "Build & serve"}
+          </button>
+        </div>
+        <p className="hint">
+          First build downloads the base model (~10&nbsp;GB) and takes a few minutes; later builds
+          reuse it.
+        </p>
+      </div>
+
+      {jobs.length === 0 && (
+        <p className="hint">No builds yet. Paste a repo URL above to start one.</p>
+      )}
+      {jobs.map((j) => (
+        <JobCard key={j.job_id} job={j} base={base} />
+      ))}
+    </div>
+  );
+}

app/src/app/v1/[...path]/route.ts

@@ -11,6 +11,11 @@ export const dynamic = "force-dynamic";
 const SERVED_MODEL = "memory-lora";
 const API_KEY = process.env.MLORA_API_KEY || "";
 
+/** True when the caller explicitly asked for the un-adapted base model. */
+function isBaseModel(model: unknown): boolean {
+  return typeof model === "string" && model.trim().toLowerCase() === "base";
+}
+
 function authOk(req: NextRequest): boolean {
   if (!API_KEY) return true;
   const auth = req.headers.get("authorization");
@@ -53,12 +58,15 @@ async function proxyPassthrough(
     );
   }
 
-  // Force the served model name so arbitrary client model strings work.
+  // Normalize the model field so arbitrary client model strings work — but
+  // NEVER collapse "base" into the served name. The upstream server selects
+  // frozen-vs-adapted from this field, so rewriting it unconditionally made the
+  // side-by-side comparison silently return the adapted model on BOTH sides.
   let body = rawBody;
   if (rawBody) {
     try {
       const j = JSON.parse(rawBody);
-      j.model = SERVED_MODEL;
+      if (!isBaseModel(j.model)) j.model = SERVED_MODEL;
       body = JSON.stringify(j);
     } catch {
       /* leave as-is */
@@ -96,7 +104,8 @@ async function handleMessages(req: NextRequest, rawBody: string): Promise<Respon
     );
   }
 
-  const oaiReq = anthropicToOpenAI(body, SERVED_MODEL);
+  const oaiReq = anthropicToOpenAI(
+    body, isBaseModel(body.model) ? "base" : SERVED_MODEL);
   const upstream = await fetch(`${target}/v1/chat/completions`, {
     method: "POST",
     headers: { "content-type": "application/json" },

data/real_code2lora/code2lora-evo/.cache/huggingface/.gitignore

@@ -1 +0,0 @@
-*

data/real_code2lora/code2lora-evo/.cache/huggingface/CACHEDIR.TAG

@@ -1,4 +0,0 @@
-Signature: 8a477f597d28d172789f06886806bc55
-# This file is a cache directory tag created by huggingface_hub.
-# For information about cache directory tags, see:
-#	https://bford.info/cachedir/

data/real_code2lora/code2lora-evo/.cache/huggingface/download/.gitattributes.metadata

@@ -1,3 +0,0 @@
-93a9e8240b9482092ecc768b804cc62bae1e875a
-bed0738c7eeb449bca98b5d2f33c89a1ee56349a
-1784887206.240134

data/real_code2lora/code2lora-evo/.cache/huggingface/download/README.md.metadata

@@ -1,3 +0,0 @@
-93a9e8240b9482092ecc768b804cc62bae1e875a
-12076d4a9bd34d0eed0b624159820e7c340ac3df
-1784887206.477736

data/real_code2lora/code2lora-evo/.cache/huggingface/download/commits/cr_test.parquet.metadata

@@ -1,3 +0,0 @@
-93a9e8240b9482092ecc768b804cc62bae1e875a
-1472bd42bed51428bf1bccbedc4c494db58bc9a6fffa72fc95f95a0613de2169
-1784887250.161078

data/real_code2lora/code2lora-evo/.cache/huggingface/download/commits/cr_val.parquet.metadata

@@ -1,3 +0,0 @@
-93a9e8240b9482092ecc768b804cc62bae1e875a
-f73959d38ad7d479a3b6715d96ae22a42ea89c4e9b1830e3f9354ebada98877f
-1784887281.485605

data/real_code2lora/code2lora-evo/.cache/huggingface/download/commits/train.parquet.metadata

@@ -1,3 +0,0 @@
-93a9e8240b9482092ecc768b804cc62bae1e875a
-6dfca1e1791fde09c1f310f2be9273b762332f28563d203a2ed7f0136cb3590b
-1784887367.7020202

data/real_code2lora/code2lora-evo/.cache/huggingface/download/qna/cr_test.parquet.metadata

@@ -1,3 +0,0 @@
-93a9e8240b9482092ecc768b804cc62bae1e875a
-fc42611e215975824582df0da76984f6a65c3d5e29002a2b205fb5437cde907b
-1784887229.2656882

data/real_code2lora/code2lora-evo/.cache/huggingface/download/qna/cr_val.parquet.metadata

@@ -1,3 +0,0 @@
-93a9e8240b9482092ecc768b804cc62bae1e875a
-3d76ca372ce366ab6f2b51765c9b5cbfbbfa33019b4e05d1dc05c25cacbea04f
-1784887263.0436091

data/real_code2lora/code2lora-evo/.cache/huggingface/download/qna/ir_test.parquet.metadata

@@ -1,3 +0,0 @@
-93a9e8240b9482092ecc768b804cc62bae1e875a
-af20fba7881a55b68efd82302b108615c6a5ae05be8e71d49c3455a7bbb124a4
-1784887233.92661

data/real_code2lora/code2lora-evo/.cache/huggingface/download/qna/ir_val.parquet.metadata

@@ -1,3 +0,0 @@
-93a9e8240b9482092ecc768b804cc62bae1e875a
-6b830ddb82490f1c7b6c44104cdcbcd1a8c9ca5f31d3c04af0421d2f0825c6a3
-1784887277.8217332

data/real_code2lora/code2lora-evo/.cache/huggingface/download/qna/train.parquet.metadata

@@ -1,3 +0,0 @@
-93a9e8240b9482092ecc768b804cc62bae1e875a
-519a092970f55750571c5223b1cfbcb0d36745a678cf41b48d4de6104c2a34e7
-1784887371.999535

data/real_code2lora/code2lora-evo/.cache/huggingface/trees/93a9e8240b9482092ecc768b804cc62bae1e875a.json

@@ -1,69 +0,0 @@
-{
- "format_version": 1,
- "files": {
-  ".gitattributes": {
-   "size": 2504,
-   "blob_id": "bed0738c7eeb449bca98b5d2f33c89a1ee56349a"
-  },
-  "README.md": {
-   "size": 1735,
-   "blob_id": "12076d4a9bd34d0eed0b624159820e7c340ac3df"
-  },
-  "commits/cr_test.parquet": {
-   "size": 68411936,
-   "blob_id": "1e2ac27253f7cf41412b4a38520e97531302d968",
-   "lfs_sha256": "1472bd42bed51428bf1bccbedc4c494db58bc9a6fffa72fc95f95a0613de2169",
-   "lfs_size": 68411936,
-   "xet_hash": "6e8f60efddd8c7caa8df21b2abec0198eb76cf6f6a455bdca73db38493870f0d"
-  },
-  "commits/cr_val.parquet": {
-   "size": 88017946,
-   "blob_id": "75709b935d623b7edb5ccd83dc0282d1440ac580",
-   "lfs_sha256": "f73959d38ad7d479a3b6715d96ae22a42ea89c4e9b1830e3f9354ebada98877f",
-   "lfs_size": 88017946,
-   "xet_hash": "471f20effb87b24b7cb2f743e483c2e8d0799df56f1d33ff029e45552e31bbfb"
-  },
-  "commits/train.parquet": {
-   "size": 612508015,
-   "blob_id": "ff02daf9b1c12e369523125b664706b7dba70169",
-   "lfs_sha256": "6dfca1e1791fde09c1f310f2be9273b762332f28563d203a2ed7f0136cb3590b",
-   "lfs_size": 612508015,
-   "xet_hash": "3d5479420b6a2cd17c9556b3235df14df1caacc9f75eac1e530a76f9a8d1d28a"
-  },
-  "qna/cr_test.parquet": {
-   "size": 16240753,
-   "blob_id": "5d5e537f81519bb5fc611cffc1d2f3e6a643403d",
-   "lfs_sha256": "fc42611e215975824582df0da76984f6a65c3d5e29002a2b205fb5437cde907b",
-   "lfs_size": 16240753,
-   "xet_hash": "8ee16b74836d3602cf10e960cba16f083e25443e7803da5f411c9af49b3b0b12"
-  },
-  "qna/cr_val.parquet": {
-   "size": 23488346,
-   "blob_id": "b979c9c6fc248831d906057a88aa7c123eede5b4",
-   "lfs_sha256": "3d76ca372ce366ab6f2b51765c9b5cbfbbfa33019b4e05d1dc05c25cacbea04f",
-   "lfs_size": 23488346,
-   "xet_hash": "a7764c559c4dda09f842dd904c11d4c7ae340c3b3d941e15fb5e6706cacec1be"
-  },
-  "qna/ir_test.parquet": {
-   "size": 22805812,
-   "blob_id": "e997fa64278ff967d0c08c22bb9c7a5571c343d5",
-   "lfs_sha256": "af20fba7881a55b68efd82302b108615c6a5ae05be8e71d49c3455a7bbb124a4",
-   "lfs_size": 22805812,
-   "xet_hash": "9c268ba844d32ab694baf09c446c689289b9e873e4de8f9b9d3cae5470380450"
-  },
-  "qna/ir_val.parquet": {
-   "size": 20154479,
-   "blob_id": "4936c3c5f14b99c30b094b696a7df5f566745ffb",
-   "lfs_sha256": "6b830ddb82490f1c7b6c44104cdcbcd1a8c9ca5f31d3c04af0421d2f0825c6a3",
-   "lfs_size": 20154479,
-   "xet_hash": "d51d246ad2ee31868d3a7d16890466349427120f47ef7fc278181d8e1e3d879d"
-  },
-  "qna/train.parquet": {
-   "size": 135352731,
-   "blob_id": "647f3e27841055c9538eec4495520d63e30f8cb6",
-   "lfs_sha256": "519a092970f55750571c5223b1cfbcb0d36745a678cf41b48d4de6104c2a34e7",
-   "lfs_size": 135352731,
-   "xet_hash": "2a1f2087a42350f3ab6f79cda6d7be87ef81f70f3b25f0ed717463812ccdb951"
-  }
- }
-}

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/.gitignore

@@ -1 +0,0 @@
-*

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/CACHEDIR.TAG

@@ -1,4 +0,0 @@
-Signature: 8a477f597d28d172789f06886806bc55
-# This file is a cache directory tag created by huggingface_hub.
-# For information about cache directory tags, see:
-#	https://bford.info/cachedir/

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/download/.gitattributes.metadata

@@ -1,3 +0,0 @@
-65b7c77eed9398fc656052d058e9af7c2d432177
-bed0738c7eeb449bca98b5d2f33c89a1ee56349a
-1784887199.592155

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/download/README.md.metadata

@@ -1,3 +0,0 @@
-65b7c77eed9398fc656052d058e9af7c2d432177
-1ee8621fd5030865aab5bef7ac72820d1778810c
-1784887199.592134

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/download/commits/cr_test.parquet.metadata

@@ -1,3 +0,0 @@
-65b7c77eed9398fc656052d058e9af7c2d432177
-2f9f2dfd78618cbb7e39c73deaed6ff9508fccd2491c8aeb4f172797d26b41aa
-1784887229.378752

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/download/commits/cr_val.parquet.metadata

@@ -1,3 +0,0 @@
-65b7c77eed9398fc656052d058e9af7c2d432177
-281812bd8d81620a83fa8bac002dbbda6c71ebbfd6a47c1df5c623be41992591
-1784887234.453962

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/download/commits/ir_test.parquet.metadata

@@ -1,3 +0,0 @@
-65b7c77eed9398fc656052d058e9af7c2d432177
-e0cffc06d2c2f5d42ad42f2ce1b6b32ec54dbde46691cc78bc1e2bd4bb5d02e2
-1784887216.458067

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/download/commits/ir_val.parquet.metadata

@@ -1,3 +0,0 @@
-65b7c77eed9398fc656052d058e9af7c2d432177
-1efa4e7298807fe75ecd81a099908a42e3378eed80cd568659c816f96ab19fa1
-1784887206.041879

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/download/commits/train.parquet.metadata

@@ -1,3 +0,0 @@
-65b7c77eed9398fc656052d058e9af7c2d432177
-e9af541dd34eaf8ec4287b06dfc1b3a5b29f90f1ce7499a51855695786d6a3d2
-1784887202.258442

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/download/qna/cr_test.parquet.metadata

@@ -1,3 +0,0 @@
-65b7c77eed9398fc656052d058e9af7c2d432177
-fc42611e215975824582df0da76984f6a65c3d5e29002a2b205fb5437cde907b
-1784887213.327064

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/download/qna/cr_val.parquet.metadata

@@ -1,3 +0,0 @@
-65b7c77eed9398fc656052d058e9af7c2d432177
-3d76ca372ce366ab6f2b51765c9b5cbfbbfa33019b4e05d1dc05c25cacbea04f
-1784887225.967221

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/download/qna/ir_test.parquet.metadata

@@ -1,3 +0,0 @@
-65b7c77eed9398fc656052d058e9af7c2d432177
-af20fba7881a55b68efd82302b108615c6a5ae05be8e71d49c3455a7bbb124a4
-1784887241.1003609

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/download/qna/ir_val.parquet.metadata

@@ -1,3 +0,0 @@
-65b7c77eed9398fc656052d058e9af7c2d432177
-6b830ddb82490f1c7b6c44104cdcbcd1a8c9ca5f31d3c04af0421d2f0825c6a3
-1784887232.856902

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/download/qna/train.parquet.metadata

@@ -1,3 +0,0 @@
-65b7c77eed9398fc656052d058e9af7c2d432177
-b2fbe5bcd7e916c8d750b654a5b0b7286578f938ed65985eca68cd2936fb5dd1
-1784887223.748132

data/real_code2lora/code2lora-static-anchor/.cache/huggingface/trees/65b7c77eed9398fc656052d058e9af7c2d432177.json

@@ -1,83 +0,0 @@
-{
- "format_version": 1,
- "files": {
-  ".gitattributes": {
-   "size": 2504,
-   "blob_id": "bed0738c7eeb449bca98b5d2f33c89a1ee56349a"
-  },
-  "README.md": {
-   "size": 1798,
-   "blob_id": "1ee8621fd5030865aab5bef7ac72820d1778810c"
-  },
-  "commits/cr_test.parquet": {
-   "size": 20979060,
-   "blob_id": "f1a4c3af961e9f98348e4d7ca9ee7fb9ecc07bbc",
-   "lfs_sha256": "2f9f2dfd78618cbb7e39c73deaed6ff9508fccd2491c8aeb4f172797d26b41aa",
-   "lfs_size": 20979060,
-   "xet_hash": "dccd2b6b33cb58024aba877ce8638a79576dbe444b902e054c45d0cc0f873f3a"
-  },
-  "commits/cr_val.parquet": {
-   "size": 26834450,
-   "blob_id": "240e8431f81a6332c9b954f4254129e6bef363e4",
-   "lfs_sha256": "281812bd8d81620a83fa8bac002dbbda6c71ebbfd6a47c1df5c623be41992591",
-   "lfs_size": 26834450,
-   "xet_hash": "bfcfcb58af8285525bc2e48a53fe34c40e8d0bb1ba0949e9f6bc8d3f1889211e"
-  },
-  "commits/ir_test.parquet": {
-   "size": 18069075,
-   "blob_id": "869388baf5ac4d1634bf743c988829540ba0318f",
-   "lfs_sha256": "e0cffc06d2c2f5d42ad42f2ce1b6b32ec54dbde46691cc78bc1e2bd4bb5d02e2",
-   "lfs_size": 18069075,
-   "xet_hash": "182e622aa94cc1f257d5e8d43b30f54e4b2497dd66dfdd3d54884d42fc27f02a"
-  },
-  "commits/ir_val.parquet": {
-   "size": 16741758,
-   "blob_id": "4a0def30c7d746a5eead7de1a977131825ccfc49",
-   "lfs_sha256": "1efa4e7298807fe75ecd81a099908a42e3378eed80cd568659c816f96ab19fa1",
-   "lfs_size": 16741758,
-   "xet_hash": "ba00255aaa6c3f1bd3e297fd8917868a79591eb66bbc2051d79626d3628df556"
-  },
-  "commits/train.parquet": {
-   "size": 1608634,
-   "blob_id": "ef5092d91d972ceafe29e6591c2cb145255cdfa9",
-   "lfs_sha256": "e9af541dd34eaf8ec4287b06dfc1b3a5b29f90f1ce7499a51855695786d6a3d2",
-   "lfs_size": 1608634,
-   "xet_hash": "45120132a92dfd70678b173eb110756e934fc6567c1347e04d3272aa148b5c1c"
-  },
-  "qna/cr_test.parquet": {
-   "size": 16240753,
-   "blob_id": "5d5e537f81519bb5fc611cffc1d2f3e6a643403d",
-   "lfs_sha256": "fc42611e215975824582df0da76984f6a65c3d5e29002a2b205fb5437cde907b",
-   "lfs_size": 16240753,
-   "xet_hash": "8ee16b74836d3602cf10e960cba16f083e25443e7803da5f411c9af49b3b0b12"
-  },
-  "qna/cr_val.parquet": {
-   "size": 23488346,
-   "blob_id": "b979c9c6fc248831d906057a88aa7c123eede5b4",
-   "lfs_sha256": "3d76ca372ce366ab6f2b51765c9b5cbfbbfa33019b4e05d1dc05c25cacbea04f",
-   "lfs_size": 23488346,
-   "xet_hash": "a7764c559c4dda09f842dd904c11d4c7ae340c3b3d941e15fb5e6706cacec1be"
-  },
-  "qna/ir_test.parquet": {
-   "size": 22805812,
-   "blob_id": "e997fa64278ff967d0c08c22bb9c7a5571c343d5",
-   "lfs_sha256": "af20fba7881a55b68efd82302b108615c6a5ae05be8e71d49c3455a7bbb124a4",
-   "lfs_size": 22805812,
-   "xet_hash": "9c268ba844d32ab694baf09c446c689289b9e873e4de8f9b9d3cae5470380450"
-  },
-  "qna/ir_val.parquet": {
-   "size": 20154479,
-   "blob_id": "4936c3c5f14b99c30b094b696a7df5f566745ffb",
-   "lfs_sha256": "6b830ddb82490f1c7b6c44104cdcbcd1a8c9ca5f31d3c04af0421d2f0825c6a3",
-   "lfs_size": 20154479,
-   "xet_hash": "d51d246ad2ee31868d3a7d16890466349427120f47ef7fc278181d8e1e3d879d"
-  },
-  "qna/train.parquet": {
-   "size": 7436646,
-   "blob_id": "b2fcda6096ba240d7aa7bd6c1d8ab8ef1ef7eb2c",
-   "lfs_sha256": "b2fbe5bcd7e916c8d750b654a5b0b7286578f938ed65985eca68cd2936fb5dd1",
-   "lfs_size": 7436646,
-   "xet_hash": "1a44ec7585045f24c0fe8a3eb42d7068e552d0bc74fc5c94af75cb66c74a1cdf"
-  }
- }
-}

data/real_code2lora/code2lora-static/.cache/huggingface/.gitignore

@@ -1 +0,0 @@
-*

data/real_code2lora/code2lora-static/.cache/huggingface/CACHEDIR.TAG

@@ -1,4 +0,0 @@
-Signature: 8a477f597d28d172789f06886806bc55
-# This file is a cache directory tag created by huggingface_hub.
-# For information about cache directory tags, see:
-#	https://bford.info/cachedir/

data/real_code2lora/code2lora-static/.cache/huggingface/download/.gitattributes.metadata

@@ -1,3 +0,0 @@
-5ebfd4fabc36b4daf5fb2b056f48ce2d67d3a9d4
-bed0738c7eeb449bca98b5d2f33c89a1ee56349a
-1784887242.232571

data/real_code2lora/code2lora-static/.cache/huggingface/download/README.md.metadata

@@ -1,3 +0,0 @@
-5ebfd4fabc36b4daf5fb2b056f48ce2d67d3a9d4
-0a390fc9888809d1fa519fee1965c070112e15dd
-1784887242.3327308

data/real_code2lora/code2lora-static/.cache/huggingface/download/qna/cr_test.parquet.metadata

@@ -1,3 +0,0 @@
-5ebfd4fabc36b4daf5fb2b056f48ce2d67d3a9d4
-234545feb01789ccfd422d0a02eea7bb43e5587e6be3e90ceca24012b7e665d7
-1784887249.4468322

data/real_code2lora/code2lora-static/.cache/huggingface/download/qna/cr_val.parquet.metadata

@@ -1,3 +0,0 @@
-5ebfd4fabc36b4daf5fb2b056f48ce2d67d3a9d4
-9c05111449e596a1817f61f3b2d728f9ae11031a425bc46e565373897db49272
-1784887248.887584

data/real_code2lora/code2lora-static/.cache/huggingface/download/qna/ir_test.parquet.metadata

@@ -1,3 +0,0 @@
-5ebfd4fabc36b4daf5fb2b056f48ce2d67d3a9d4
-7e6261be02c00201b7ed61950cebbeb1297f19f1de92ede70a9e63943747476a
-1784887255.800129

data/real_code2lora/code2lora-static/.cache/huggingface/download/qna/ir_val.parquet.metadata

@@ -1,3 +0,0 @@
-5ebfd4fabc36b4daf5fb2b056f48ce2d67d3a9d4
-2a1ba08fee4af69d321c8278493a5c3256ea96861c7f35d785e7c90be4657ec6
-1784887257.502165

data/real_code2lora/code2lora-static/.cache/huggingface/download/qna/train.parquet.metadata

@@ -1,3 +0,0 @@
-5ebfd4fabc36b4daf5fb2b056f48ce2d67d3a9d4
-d5657224e62418a77187285fe55a650ebad1a264bb9edbdc20c7d1e132c9b523
-1784887281.037739

data/real_code2lora/code2lora-static/.cache/huggingface/download/repos/cr_test.parquet.metadata

@@ -1,3 +0,0 @@
-5ebfd4fabc36b4daf5fb2b056f48ce2d67d3a9d4
-0d597c8a2462fd407394db40ab9025dc4f4ccc0697a2d4490fdd2a1eab91578f
-1784887243.962303

data/real_code2lora/code2lora-static/.cache/huggingface/download/repos/cr_val.parquet.metadata

@@ -1,3 +0,0 @@
-5ebfd4fabc36b4daf5fb2b056f48ce2d67d3a9d4
-8ac1c2477884783356be16f4e2ec5c94c60823df8befb4a2d80813f40c84b794
-1784887244.5128582

data/real_code2lora/code2lora-static/.cache/huggingface/download/repos/train.parquet.metadata

@@ -1,3 +0,0 @@
-5ebfd4fabc36b4daf5fb2b056f48ce2d67d3a9d4
-ad382500d5bb420e0c8064e7f79d8a26fc12aa8e5b209c95b145a0b4a18dd8ba
-1784887245.4021702