Add score-vs-latency chart; fix pipeline_tag to token-classification

README.md

@@ -11,7 +11,7 @@ tags:
   - bitnet
   - gguf
 library_name: gguf
-pipeline_tag: text2text-generation
+pipeline_tag: token-classification
 ---
 
 # SlothE-T 25M — Ternary Zhuyin → Traditional Chinese IME model
@@ -64,6 +64,25 @@ window, filtered to be **absent from the training corpus**), then g2pW-labeled.
 > 25M ternary model **beats the previously-shipped 12M int8 on 免選字 and
 > homophone-hard** while being smaller and faster.
 
+### Quality vs. latency
+
+![Held-out 免選字 vs. on-device latency](score_vs_latency.svg)
+
+Latency is per-6-syllable decode on a **BOOX (Snapdragon 662, no-dotprod ARM)**.
+4M and 12M are measured (ORT int8); the **25M ternary** point is *projected* from
+measured-shape TQ2_0 / I2_S kernel benchmarks (the ternary win is diluted by the
+int8 8342-way head, which ternary can't accelerate). The takeaway: the 25M ternary
+is **Pareto-dominant** — roughly the 4M's latency (~9 ms) but **+6 免選字**, and
+both faster *and* more accurate than the 12M.
+
+| model | params | 免選字 | homophone | toneless | latency (BOOX) |
+|---|---|---|---|---|---|
+| 4M int8 | 4M | 70 % | 83 % | 74 % | 9.1 ms |
+| 12M int8 | 12M | 72 % | 82 % | 79 % | 13.3 ms |
+| **25M ternary** | **25M** | **76 %** | **86 %** | **77 %** | **~9 ms\*** |
+
+<sub>\* projected, not full end-to-end measured. See the [experiment notes](REPRODUCE.md).</sub>
+
 ## Training recipe (teacher-free)
 
 - **Direct cross-entropy** on a g2pW-labeled zh-TW corpus, with **label smoothing 0.1**.

make_plot.py

@@ -0,0 +1,81 @@
+#!/usr/bin/env python3
+"""Score-vs-latency scatter for the model card. Self-contained SVG, no deps."""
+
+# honest held-out (500 fresh zh-TW sentences) + BOOX (SD662) per-6-syllable decode ms
+# latency: 4M/12M measured (ORT int8); 25M ternary projected from measured-shape
+#          bitnet TQ2_0/I2_S kernel benchmarks (marked with *).
+MODELS = [
+    # name,           params, latency_ms, proj, mianxuan, homophone, toneless, color
+    ("4M int8",       "4M",   9.06, False, 70, 83, 74, "#6b7280"),
+    ("12M int8",      "12M",  13.3, False, 72, 82, 79, "#2563eb"),
+    ("25M ternary",   "25M",  9.0,  True,  76, 86, 77, "#dc2626"),
+]
+
+W, H = 680, 440
+L, R, T, B = 74, 28, 46, 66          # margins
+PX0, PX1 = L, W - R
+PY0, PY1 = T, H - B
+XMIN, XMAX = 8.0, 14.5
+YMIN, YMAX = 66.0, 80.0
+
+def x(v): return PX0 + (v - XMIN) / (XMAX - XMIN) * (PX1 - PX0)
+def y(v): return PY1 - (v - YMIN) / (YMAX - YMIN) * (PY1 - PY0)
+
+s = []
+s.append(f'<svg xmlns="http://www.w3.org/2000/svg" width="{W}" height="{H}" '
+         f'viewBox="0 0 {W} {H}" font-family="-apple-system,Segoe UI,Roboto,sans-serif">')
+# always-light background card so it reads on both HF themes
+s.append(f'<rect x="0" y="0" width="{W}" height="{H}" rx="10" fill="#ffffff"/>')
+s.append(f'<text x="{L}" y="26" font-size="16" font-weight="700" fill="#111827">'
+         f'Held-out quality vs. on-device latency (BOOX SD662)</text>')
+
+# gridlines + y ticks (免選字 %)
+for yv in range(66, 81, 2):
+    yy = y(yv)
+    s.append(f'<line x1="{PX0}" y1="{yy:.1f}" x2="{PX1}" y2="{yy:.1f}" stroke="#e5e7eb"/>')
+    s.append(f'<text x="{PX0-10}" y="{yy+4:.1f}" font-size="12" fill="#6b7280" '
+             f'text-anchor="end">{yv}</text>')
+# x ticks (ms)
+for xv in range(8, 15, 1):
+    xx = x(xv)
+    s.append(f'<line x1="{xx:.1f}" y1="{PY0}" x2="{xx:.1f}" y2="{PY1}" stroke="#f3f4f6"/>')
+    s.append(f'<text x="{xx:.1f}" y="{PY1+20}" font-size="12" fill="#6b7280" '
+             f'text-anchor="middle">{xv}</text>')
+# axis labels
+s.append(f'<text x="{(PX0+PX1)/2:.0f}" y="{H-14}" font-size="13" fill="#374151" '
+         f'text-anchor="middle">latency — ms / 6-syllable decode  (lower = faster)</text>')
+s.append(f'<text x="18" y="{(PY0+PY1)/2:.0f}" font-size="13" fill="#374151" '
+         f'text-anchor="middle" transform="rotate(-90 18 {(PY0+PY1)/2:.0f})">'
+         f'免選字 (whole-sentence) %  (higher = better)</text>')
+# axis frame
+s.append(f'<rect x="{PX0}" y="{PY0}" width="{PX1-PX0}" height="{PY1-PY0}" '
+         f'fill="none" stroke="#9ca3af"/>')
+
+# Pareto arrow: 4M -> 25M ternary (up, same latency), the headline story
+s.append(f'<line x1="{x(9.06):.1f}" y1="{y(70)-10:.1f}" x2="{x(9.0):.1f}" '
+         f'y2="{y(76)+12:.1f}" stroke="#dc2626" stroke-width="1.5" '
+         f'stroke-dasharray="4 3" opacity="0.6"/>')
+
+# points
+for name, tag, lat, proj, mx, ho, tl, col in MODELS:
+    cx, cy = x(lat), y(mx)
+    s.append(f'<circle cx="{cx:.1f}" cy="{cy:.1f}" r="8" fill="{col}" '
+             f'stroke="#ffffff" stroke-width="2"/>')
+    star = "*" if proj else ""
+    # two-line label fully above the point (name higher, stats just above point),
+    # except 4M which sits low so its label goes below the point.
+    if name == "4M int8":
+        name_y, sub_y = cy + 24, cy + 39
+    else:
+        name_y, sub_y = cy - 30, cy - 14
+    s.append(f'<text x="{cx:.1f}" y="{name_y:.1f}" font-size="13" font-weight="700" '
+             f'fill="{col}" text-anchor="middle">{name}{star}</text>')
+    s.append(f'<text x="{cx:.1f}" y="{sub_y:.1f}" font-size="11" fill="#6b7280" '
+             f'text-anchor="middle">免{mx} · 同{ho} · 無{tl}</text>')
+
+# projection note under the title (avoids the crowded bottom axis area)
+s.append(f'<text x="{L}" y="42" font-size="11" fill="#9ca3af">'
+         f'* 25M ternary latency projected from measured-shape TQ2_0 / I2_S kernels</text>')
+s.append('</svg>')
+open("score_vs_latency.svg", "w").write("\n".join(s))
+print("wrote score_vs_latency.svg")