File size: 16,177 Bytes
c02657c
 
 
 
 
 
 
 
 
 
 
 
 
73f7e6f
c02657c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fb6c876
 
 
 
 
 
 
 
 
c02657c
 
 
 
 
 
 
 
 
 
73f7e6f
c02657c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
73f7e6f
c02657c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
73f7e6f
c02657c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
73f7e6f
c02657c
 
 
 
 
 
73f7e6f
c02657c
 
 
 
 
 
73f7e6f
c02657c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
"""Render the dataset-card charts from out/stats/*.json into out/assets/*.png.

Measured charts use real numbers produced by build.py / train/*.py. Projection charts
are computed from published scaling laws and are labelled as projections.
"""
import json
import math
import os

import matplotlib

matplotlib.use("Agg")
import matplotlib.pyplot as plt  # noqa: E402
import matplotlib.ticker  # noqa: E402
import numpy as np  # noqa: E402

OUT = os.environ.get("OPUSDB_OUT", "out")
ST = os.path.join(OUT, "stats")
AS = os.path.join(OUT, "assets")
os.makedirs(AS, exist_ok=True)

# Validated categorical palette (fixed order), light surface.
C = ["#2a78d6", "#eb6834", "#1baf7a", "#eda100", "#e87ba4", "#008300", "#6250d6", "#e34948"]
SURFACE, INK, INK2, GRID = "#fcfcfb", "#0b0b0b", "#52514e", "#e6e5e0"

plt.rcParams.update({
    "figure.facecolor": SURFACE, "axes.facecolor": SURFACE, "savefig.facecolor": SURFACE,
    "axes.edgecolor": GRID, "axes.labelcolor": INK2, "xtick.color": INK2, "ytick.color": INK2,
    "text.color": INK, "axes.grid": True, "grid.color": GRID, "grid.linewidth": 0.8,
    "axes.spines.top": False, "axes.spines.right": False, "font.size": 10.5,
    "axes.titlesize": 13, "axes.titleweight": "bold", "axes.titlelocation": "left", "figure.dpi": 150,
})


def load(name):
    p = os.path.join(ST, name + ".json")
    return json.load(open(p)) if os.path.exists(p) else None


def save(fig, name):
    fig.tight_layout()
    fig.savefig(os.path.join(AS, name), bbox_inches="tight")
    plt.close(fig)
    print("chart", name)


def human(n):
    for unit, k in (("B", 1e9), ("M", 1e6), ("K", 1e3)):
        if n >= k:
            return f"{n / k:.1f}{unit}" if n / k < 100 else f"{n / k:.0f}{unit}"
    return str(int(n))


# ------------------------------------------------------------------ 1. token composition per config
def composition():
    sft, pref, pre = load("sft_sources"), load("preference_sources"), load("pretrain_tokens")
    tools = load("tools_sources")
    panels = [("pretrain", {k: {"tokens": v} for k, v in (pre or {}).items()}), ("sft", sft or {}),
              ("preference", pref or {}), ("tools", tools or {})]
    panels = [p for p in panels if p[1]]
    fig, axes = plt.subplots(2, 2, figsize=(15, 9.5))
    axes = axes.flatten()
    for ax in axes[len(panels):]:
        ax.axis("off")
    for ax, (title, data) in zip(axes, panels):
        items = sorted(data.items(), key=lambda kv: -kv[1]["tokens"])
        names = [k for k, _ in items]
        vals = [v["tokens"] for _, v in items]
        y = np.arange(len(names))[::-1]
        ax.barh(y, vals, color=C[0], height=0.62)
        for yi, v in zip(y, vals):
            ax.text(v, yi, " " + human(v), va="center", fontsize=9, color=INK2)
        ax.set_yticks(y, names)
        ax.set_title(f"{title}: {human(sum(vals))} tokens")
        ax.set_xlabel("tokens (SmolLM2 tokenizer)")
        ax.xaxis.set_major_formatter(matplotlib.ticker.FuncFormatter(lambda v, _: human(v) if v else "0"))
        ax.grid(axis="y", visible=False)
        ax.set_xlim(0, max(vals or [1]) * 1.25)
    fig.suptitle("OPUS-5.5-DATABASE — token composition by config and source (measured)", x=0.01, ha="left",
                 fontweight="bold", fontsize=14)
    save(fig, "composition.png")


# ------------------------------------------------------------------ 2. SFT domain + language mix
def sft_mix():
    sft = load("sft_sources")
    if not sft:
        return
    domains = {}
    from collections import Counter
    langs = Counter()
    dom_map = {"smol-magpie-ultra": "general chat", "smol-constraints": "instruction following",
               "smol-rewrite": "rewriting", "smol-summarize": "summarization", "oasst2": "human-written chat",
               "metamathqa": "math", "openmathinstruct2": "math", "gsm8k": "math", "magicoder-oss": "code",
               "codealpaca": "code", "opus-synth": "verified synthetic (EN/TR)"}
    for k, v in sft.items():
        d = dom_map.get(k, k)
        domains[d] = domains.get(d, 0) + v["rows"]
        for kk, vv in v.items():
            if kk.startswith("lang_"):
                langs[kk[5:]] += vv
    fig, (a1, a2) = plt.subplots(1, 2, figsize=(14, 4.8), gridspec_kw={"width_ratios": [1.5, 1]})
    items = sorted(domains.items(), key=lambda kv: -kv[1])
    y = np.arange(len(items))[::-1]
    tot = sum(v for _, v in items)
    a1.barh(y, [v for _, v in items], color=C[0], height=0.62)
    for yi, (_, v) in zip(y, items):
        a1.text(v, yi, f" {v:,} ({100 * v / tot:.1f}%)", va="center", fontsize=9, color=INK2)
    a1.set_yticks(y, [k for k, _ in items])
    a1.set_xlim(0, items[0][1] * 1.35)
    a1.grid(axis="y", visible=False)
    a1.set_title(f"SFT conversations by capability ({tot:,} total)")
    top = langs.most_common(6)
    other = sum(langs.values()) - sum(v for _, v in top)
    labels = [k for k, _ in top] + (["all remaining"] if other else [])
    vals = [v for _, v in top] + ([other] if other else [])
    y2 = np.arange(len(labels))[::-1]
    a2.barh(y2, vals, color=C[2], height=0.62)
    for yi, v in zip(y2, vals):
        a2.text(v, yi, f" {v:,}", va="center", fontsize=9, color=INK2)
    a2.set_yticks(y2, labels)
    a2.set_xscale("log")
    a2.set_xlim(min(vals) / 2, max(vals) * 8)
    a2.grid(axis="y", visible=False)
    a2.set_title("SFT conversations by language (log scale)")
    save(fig, "sft_mix.png")


# ------------------------------------------------------------------ 3. length distribution
def lengths():
    sl = load("sft_lengths")
    pl = load("pretrain_lengths")
    fig, (a1, a2) = plt.subplots(1, 2, figsize=(14, 4.6))
    if sl:
        bins = np.logspace(1, math.log10(8192), 50)
        groups = {"chat & instruction": ("smol-magpie-ultra", "smol-constraints", "smol-rewrite", "smol-summarize", "oasst2"),
                  "math": ("metamathqa", "openmathinstruct2", "gsm8k"),
                  "code": ("magicoder-oss", "codealpaca"), "verified synthetic": ("opus-synth",)}
        arr, src = np.array(sl["num_tokens"]), np.array(sl["source"])
        for i, (g, srcs) in enumerate(groups.items()):
            vals = arr[np.isin(src, srcs)]
            a1.hist(vals, bins=bins, histtype="step", linewidth=2, color=C[i], label=f"{g} (median {int(np.median(vals))})")
        a1.set_xscale("log")
        a1.set_xlabel("tokens per conversation")
        a1.set_ylabel("conversations")
        a1.set_title("SFT conversation length")
        a1.legend(frameon=False, fontsize=9)
    if pl:
        bins = np.logspace(1.5, 5, 50)
        for i, (k, v) in enumerate(pl.items()):
            a2.hist(v, bins=bins, histtype="step", linewidth=2, color=C[i], label=f"{k} (median {int(np.median(v))})")
        a2.set_xscale("log")
        a2.set_xlabel("tokens per document")
        a2.set_ylabel("documents (sample)")
        a2.set_title("Pretraining document length")
        a2.legend(frameon=False, fontsize=9)
    save(fig, "lengths.png")


# ------------------------------------------------------------------ 4. filtering funnel
def funnel():
    parts = [("sft", load("sft_funnel")), ("preference", load("preference_funnel")), ("pretrain", load("pretrain_funnel"))]
    rows = []
    for cfg, f in parts:
        if not f:
            continue
        for src, c in f.items():
            if not isinstance(c, dict):
                continue
            read = c.get("read", 0)
            kept = c.get("kept", 0)
            q = sum(v for k, v in c.items() if k.startswith(("quality", "bad_", "empty", "boiler", "low_", "incomplete",
                                                              "no_margin", "identical", "argilla")))
            dup = c.get("exact_duplicate", 0) + c.get("near_duplicate", 0)
            con = c.get("benchmark_contaminated", 0)
            cap = c.get("over_cap", 0)
            rows.append((f"{cfg}/{src}", read, kept, q, dup, con, cap))
    if not rows:
        return
    fig, ax = plt.subplots(figsize=(13, 0.42 * len(rows) + 1.6))
    y = np.arange(len(rows))[::-1]
    labels = ["kept", "quality / format filter", "duplicate (exact + MinHash)", "benchmark-contaminated", "subsampled (cap)"]
    cols = [C[2], C[3], C[0], C[7], "#c9c8c2"]
    for yi, r in zip(y, rows):
        left = 0
        read = max(1, r[1])
        for j, v in enumerate([r[2], r[3], r[4], r[5], r[6]]):
            w = 100 * v / read
            ax.barh(yi, w, left=left, color=cols[j], height=0.62, edgecolor=SURFACE, linewidth=1.5,
                    label=labels[j] if yi == y[0] else None)
            left += w
        ax.text(101, yi, f"{r[2]:,} / {r[1]:,}", va="center", fontsize=8.5, color=INK2)
    ax.set_yticks(y, [r[0] for r in rows], fontsize=9)
    ax.set_xlim(0, 125)
    ax.set_xticks([0, 20, 40, 60, 80, 100])
    ax.set_xlabel("% of rows read")
    ax.grid(axis="y", visible=False)
    ax.legend(ncol=5, frameon=False, fontsize=9, loc="upper left", bbox_to_anchor=(0, -0.08))
    ax.set_title("Filtering funnel per source (measured): kept rows / rows read")
    save(fig, "funnel.png")


# ------------------------------------------------------------------ 5. measured scaling ladder
def ladder():
    lad = load("ladder")
    if not lad or not lad.get("runs"):
        return
    runs = sorted(lad["runs"], key=lambda r: r["total_params"])
    fig, (a1, a2) = plt.subplots(1, 2, figsize=(14, 4.8))
    for i, r in enumerate(runs):
        xs = [p["tokens"] for p in r["curve"]]
        ys = [p["val_loss"] for p in r["curve"]]
        a1.plot(xs, ys, color=C[i], linewidth=2, label=f"{human(r['total_params'])} params")
    a1.set_xscale("log")
    a1.set_xlabel("training tokens")
    a1.set_ylabel("validation loss (nats/token)")
    a1.set_title("Measured: validation loss during training")
    a1.legend(frameon=False, fontsize=9)
    N = np.array([r["total_params"] for r in runs], float)
    L = np.array([r["val_loss"] for r in runs])
    a2.plot(N, L, "o", color=C[0], markersize=9, label="measured (final val loss)")
    fit = lad.get("fit")
    if fit:
        xs = np.logspace(math.log10(N.min() / 2), math.log10(N.max() * 30), 100)
        a2.plot(xs, fit["E"] + fit["A"] / xs ** fit["alpha"], "--", color=C[1], linewidth=2,
                label=f"fit L = {fit['E']:.2f} + {fit['A']:.3g}/N^{fit['alpha']:.2f}")
    for n, l in zip(N, L):
        a2.annotate(f"{l:.2f}", (n, l), textcoords="offset points", xytext=(6, 6), fontsize=9, color=INK2)
    a2.set_xscale("log")
    a2.set_xlabel("model parameters (N)")
    a2.set_ylabel("final validation loss")
    a2.set_title("Measured: loss vs model size on OPUS pretrain")
    a2.legend(frameon=False, fontsize=9)
    save(fig, "ladder.png")


# ------------------------------------------------------------------ 6. projections (published scaling laws)
SIZES = [0.135, 0.5, 1, 4, 8, 14, 27, 70]  # billions


def chinchilla(N, D, E=1.69, A=406.4, B=410.7, a=0.34, b=0.28):
    return E + A / N ** a + B / D ** b


def effective_data(U, D, r_star=15.4):
    """Muennighoff et al. (2023) data-constrained scaling: value of repeated tokens decays."""
    if D <= U:
        return D
    R = D / U - 1
    return U + U * r_star * (1 - math.exp(-R / r_star))


def projections():
    pre = load("pretrain_tokens")
    sft = load("sft_sources")
    tools = load("tools_sources") or {}
    if not pre:
        return
    U = sum(pre.values())
    fig, (a1, a2) = plt.subplots(1, 2, figsize=(14, 5))
    Ns = np.array(SIZES) * 1e9
    ideal = [chinchilla(n, 20 * n) for n in Ns]
    opus = [chinchilla(n, effective_data(U, 20 * n)) for n in Ns]
    opus1 = [chinchilla(n, U) for n in Ns]
    a1.plot(Ns, ideal, "-o", color=C[0], lw=2, ms=6, label="unlimited unique data, 20 tok/param (Chinchilla optimum)")
    a1.plot(Ns, opus, "-o", color=C[1], lw=2, ms=6, label=f"OPUS pretrain only ({human(U)} unique), repeated to 20 tok/param")
    a1.plot(Ns, opus1, "--", color=C[6], lw=2, label="OPUS pretrain only, single epoch")
    a1.set_xscale("log")
    a1.set_xticks(Ns, [f"{s:g}B" for s in SIZES], rotation=45)
    a1.minorticks_off()
    a1.set_ylabel("projected loss (Chinchilla units, nats)")
    a1.set_title("PROJECTION: pretraining from scratch")
    a1.legend(frameon=False, fontsize=8.5)
    # compute to fine-tune on SFT config (2 epochs) — arithmetic, not a measurement
    if sft:
        T = sum(v["tokens"] for v in sft.values()) + sum(v["tokens"] for v in tools.values())
        flops = 6 * Ns * T * 2
        h100 = flops / (989e12 * 0.40) / 3600
        lora = h100 / 3   # LoRA skips weight-gradient matmuls for frozen weights (~2/3 of backward)
        a2.plot(Ns, h100, "-o", color=C[0], lw=2, ms=6, label="full fine-tune")
        a2.plot(Ns, lora, "-o", color=C[2], lw=2, ms=6, label="LoRA (approx.)")
        for n, h in zip(Ns, h100):
            if n / 1e9 in (0.5, 4, 8, 14, 27, 70):
                a2.annotate(f"{h:,.1f} h", (n, h), textcoords="offset points", xytext=(-10, 8), fontsize=9, color=INK2)
        a2.set_xscale("log")
        a2.set_yscale("log")
        a2.set_xticks(Ns, [f"{s:g}B" for s in SIZES], rotation=45)
        a2.minorticks_off()
        a2.set_ylabel("H100 GPU-hours (40% MFU)")
        a2.set_title(f"ESTIMATE: 2 epochs on sft + tools ({human(T)} tokens)")
        a2.legend(frameon=False, fontsize=9)
    save(fig, "projections.png")
    out = {"unique_pretrain_tokens": U,
           "rows": [{"params_b": s, "chinchilla_optimal_tokens_b": 20 * s, "epochs_needed_on_opus": 20 * s * 1e9 / U,
                     "loss_unlimited": ideal[i], "loss_opus_repeated": opus[i], "loss_opus_1epoch": opus1[i]}
                    for i, s in enumerate(SIZES)]}
    json.dump(out, open(os.path.join(ST, "projections.json"), "w"), indent=2)


# ------------------------------------------------------------------ 7. measured model evals on synth_eval
def evals():
    ev = load("evals")
    if not ev:
        return
    base = [r for r in ev if r.get("kind") == "baseline"]
    tuned = [r for r in ev if r.get("kind") == "opus-sft"]
    fig, (a1, a2) = plt.subplots(1, 2, figsize=(14, 4.8), gridspec_kw={"width_ratios": [1.2, 1]})
    if base:
        base = sorted(base, key=lambda r: r["params_b"])
        x = np.arange(len(base))
        a1.bar(x - 0.2, [100 * r["by_language"]["en"] for r in base], 0.38, color=C[0], label="English items")
        a1.bar(x + 0.2, [100 * r["by_language"]["tr"] for r in base], 0.38, color=C[1], label="Turkish items")
        for xi, r in zip(x, base):
            a1.text(xi, 100 * max(r["by_language"].values()) + 2, f"{100 * r['accuracy']:.0f}%", ha="center",
                    fontsize=9, color=INK)
        a1.set_xticks(x, [f"{r['label']}\n{r['params_b']:g}B" for r in base], fontsize=9)
        a1.set_ylim(0, 100)
        a1.set_ylabel("exact-match accuracy (%)")
        a1.set_title(f"Measured: off-the-shelf models on OPUS synth_eval (n={base[0]['n']})")
        a1.grid(axis="x", visible=False)
        a1.legend(frameon=False, fontsize=9)
    if tuned:
        pairs = []
        for t in tuned:
            b = next((r for r in ev if r["model"] == t["base_model"]), None)
            if b:
                pairs.append((b, t))
        fams = sorted(pairs[0][1]["by_family"]) if pairs else []
        if pairs:
            b, t = pairs[0]
            y = np.arange(len(fams) + 1)[::-1]
            bv = [100 * b["accuracy"]] + [100 * b["by_family"].get(f, 0) for f in fams]
            tv = [100 * t["accuracy"]] + [100 * t["by_family"].get(f, 0) for f in fams]
            a2.barh(y + 0.2, bv, 0.38, color="#c9c8c2", label=f"{b['label']} (before)")
            a2.barh(y - 0.2, tv, 0.38, color=C[2], label=f"after OPUS-SFT ({t['train_examples']:,} ex.)")
            a2.set_yticks(y, ["OVERALL"] + fams, fontsize=9)
            a2.set_xlim(0, 100)
            a2.set_xlabel("exact-match accuracy (%)")
            a2.set_title("Measured: fine-tuning a small model on OPUS-SFT")
            a2.grid(axis="y", visible=False)
            a2.legend(frameon=False, fontsize=9, loc="lower right")
    else:
        a2.axis("off")
    save(fig, "evals.png")


if __name__ == "__main__":
    for f in (composition, sft_mix, lengths, funnel, ladder, projections, evals):
        f()