Maggio33 commited on
Commit
c1cee49
·
verified ·
1 Parent(s): fe60d85

Slayer1B 1.04 in progress: adoption 10010 on 8xL40S, mid-run eval, 1.04-checkpoints archive

Browse files
Files changed (1) hide show
  1. index.html +83 -7
index.html CHANGED
@@ -24,7 +24,7 @@
24
  .card a { color:var(--acc3); text-decoration:none; }
25
  .card a:hover { text-decoration:underline; }
26
  .tag { display:inline-block; font-size:.78rem; padding:2px 8px; border-radius:6px; margin-left:6px; vertical-align:middle; }
27
- .tag.dec { background:#2b2140; color:#c7b2ff; } .tag.const { background:#1d3326; color:#8fe0ad; }
28
  table { width:100%; border-collapse:collapse; background:var(--card); border:1px solid var(--line); border-radius:14px; overflow:hidden; }
29
  th, td { padding:10px 12px; text-align:right; border-bottom:1px solid var(--line); font-variant-numeric:tabular-nums; }
30
  th:first-child, td:first-child { text-align:left; }
@@ -99,6 +99,15 @@
99
  <a href="https://track.fabryka.ai/run/1426c0a8-2f07-5a03-80fc-92f7e922aa4b">checkpoint benchmarks</a> ·
100
  <a href="https://huggingface.co/SlayerLab/Slayer1B-1.03-checkpoints">native checkpoint archive — current files</a>
101
  </div>
 
 
 
 
 
 
 
 
 
102
  </div>
103
  </section>
104
 
@@ -141,6 +150,33 @@
141
  <p class="note">Measured values only, no fits: one checkpoint every 150 steps. Historical data: <a href="https://huggingface.co/spaces/SlayerLab/Slayer1B/resolve/b67cc3f1c3a56ea16f2ae92f44023e70750e371d/scores-history.csv">scores-history.csv</a>.</p>
142
  </section>
143
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
144
  <section>
145
  <h2>How we do it</h2>
146
  <ol class="steps">
@@ -164,12 +200,14 @@
164
  <section>
165
  <h2>What's next</h2>
166
  <p><b>Corpus continuation, 1.04–1.06:</b> the global trajectory runs from native step 7150 toward the target 13848.
167
- Checkpoint milestones are 10010 (1.04), 12870 (1.05) and 13848 (1.06).
168
  <a href="https://track.fabryka.ai/run/f056bb5c-f1da-56d5-9776-cc33063f10a5">Follow the live continuation</a>.</p>
169
  <p class="note">The linked runs belong to the author's Track account. The completed 1.03 run includes its resumed steps;
170
  corpus continuation and benchmark history are separate records of the same model lineage.</p>
171
- <p><b>At the end of the line:</b> one final learning-rate decay, then extending the context window in stages —
172
- first to 32k tokens, with up to 256k as the target, each stage checked with long-context tests.</p>
 
 
173
  <p class="note">These are base models: they predict the continuation of a text, they were not tuned for chat or for
174
  following instructions, and they can be wrong.</p>
175
  </section>
@@ -226,6 +264,15 @@
226
  <a href="https://track.fabryka.ai/run/1426c0a8-2f07-5a03-80fc-92f7e922aa4b">benchmarki checkpointów</a> ·
227
  <a href="https://huggingface.co/SlayerLab/Slayer1B-1.03-checkpoints">archiwum checkpointów natywnych — aktualne pliki</a>
228
  </div>
 
 
 
 
 
 
 
 
 
229
  </div>
230
  </section>
231
 
@@ -268,6 +315,33 @@
268
  <p class="note">Same pomiary, bez dopasowań: punkt kontrolny co 150 kroków. Historyczne dane: <a href="https://huggingface.co/spaces/SlayerLab/Slayer1B/resolve/b67cc3f1c3a56ea16f2ae92f44023e70750e371d/scores-history.csv">scores-history.csv</a>.</p>
269
  </section>
270
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
271
  <section>
272
  <h2>Jak to robimy</h2>
273
  <ol class="steps">
@@ -291,12 +365,14 @@
291
  <section>
292
  <h2>Co dalej</h2>
293
  <p><b>Kontynuacja korpusu, 1.04–1.06:</b> globalna trajektoria prowadzi od kroku native 7150 do celu 13848.
294
- Checkpointy wersji przypadają na 10010 (1.04), 12870 (1.05) i 13848 (1.06).
295
  <a href="https://track.fabryka.ai/run/f056bb5c-f1da-56d5-9776-cc33063f10a5">Postęp kontynuacji na żywo</a>.</p>
296
  <p class="note">Powiązane przebiegi są na koncie autora w Tracku. Zakończony run 1.03 obejmuje wznowione kroki;
297
  dalszy trening i historia benchmarków to osobne zapisy tej samej linii modelu.</p>
298
- <p><b>Na końcu linii:</b> jedno końcowe wygaszanie tempa uczenia, a potem etapami wydłużanie okna kontekstu —
299
- najpierw do 32k tokenów, docelowo do 256k, każdy etap sprawdzany testami długiego kontekstu.</p>
 
 
300
  <p class="note">To modele bazowe: przewidują dalszy ciąg tekstu, nie były dostrajane do rozmowy ani wykonywania poleceń
301
  i mogą się mylić.</p>
302
  </section>
 
24
  .card a { color:var(--acc3); text-decoration:none; }
25
  .card a:hover { text-decoration:underline; }
26
  .tag { display:inline-block; font-size:.78rem; padding:2px 8px; border-radius:6px; margin-left:6px; vertical-align:middle; }
27
+ .tag.const { background:#1d3326; color:#8fe0ad; } .tag.live { background:#3a2c17; color:#e8c37a; }
28
  table { width:100%; border-collapse:collapse; background:var(--card); border:1px solid var(--line); border-radius:14px; overflow:hidden; }
29
  th, td { padding:10px 12px; text-align:right; border-bottom:1px solid var(--line); font-variant-numeric:tabular-nums; }
30
  th:first-child, td:first-child { text-align:left; }
 
99
  <a href="https://track.fabryka.ai/run/1426c0a8-2f07-5a03-80fc-92f7e922aa4b">checkpoint benchmarks</a> ·
100
  <a href="https://huggingface.co/SlayerLab/Slayer1B-1.03-checkpoints">native checkpoint archive — current files</a>
101
  </div>
102
+ <div class="card card-status">
103
+ <h3>Slayer1B 1.04 <span class="tag live">training in progress</span></h3>
104
+ <div class="meta">corpus continuation 7150 → 10010 · target ~10.5 B tokens</div>
105
+ <p>Started on 4×L40S, adopted onto 8×L40S at the 10010 milestone after independent SHA-256 verification
106
+ of the checkpoint (three sources) and an NCCL/tempo gate (~100k tokens/s synthetic).</p>
107
+ <p class="note">Milestone 10010 SHA-256: <code class="sha">04438959d7b317d1a8d972b264e9610a06fd8addde7349c6e9f458ee56dc1e51</code></p>
108
+ <a href="https://track.fabryka.ai/run/f056bb5c-f1da-56d5-9776-cc33063f10a5">live continuation run</a> ·
109
+ <a href="https://huggingface.co/SlayerLab/Slayer1B-1.04-checkpoints">native checkpoint archive — current files</a>
110
+ </div>
111
  </div>
112
  </section>
113
 
 
150
  <p class="note">Measured values only, no fits: one checkpoint every 150 steps. Historical data: <a href="https://huggingface.co/spaces/SlayerLab/Slayer1B/resolve/b67cc3f1c3a56ea16f2ae92f44023e70750e371d/scores-history.csv">scores-history.csv</a>.</p>
151
  </section>
152
 
153
+ <section>
154
+ <h2>Mid-run evaluation: 1.04 continuation</h2>
155
+ <div class="wrap">
156
+ <table>
157
+ <tr><th>test</th><th>step 7200</th><th>step 7250</th></tr>
158
+ <tr><td>BLiMP (EN, grammar), %</td><td>74.63</td><td class="best">74.70</td></tr>
159
+ <tr><td>ARC-Easy (EN), %</td><td>49.75</td><td class="best">49.87</td></tr>
160
+ <tr><td>WikiText-2, byte perplexity ↓</td><td class="best">2.3439</td><td>2.3456</td></tr>
161
+ <tr><td>WikiText-2, token perplexity ↓</td><td class="best">33.36</td><td>33.45</td></tr>
162
+ <tr><td>MultiBLiMP (PL, grammar), %</td><td>98.41</td><td class="best">98.47</td></tr>
163
+ <tr><td>ARC-Easy-PL, %</td><td class="best">43.10</td><td>42.55</td></tr>
164
+ </table>
165
+ </div>
166
+ <p class="note">Preliminary mid-run points from the <a href="https://track.fabryka.ai/run/f056bb5c-f1da-56d5-9776-cc33063f10a5">live 1.04 continuation</a>,
167
+ measured automatically during training (lm-eval harness, one instrument). Not final results and not directly comparable
168
+ with the offline table above (different procedure). Both points share one environment, so 7200 → 7250 is a clean
169
+ within-run comparison; at 50 steps apart the changes are small and within run noise.</p>
170
+ </section>
171
+
172
+ <section>
173
+ <h2>Live charts: 1.04 continuation</h2>
174
+ <figure><svg viewBox="0 0 760 250" role="img" style="width:100%;height:auto;background:transparent" xmlns="http://www.w3.org/2000/svg"><line x1="40" y1="210" x2="720" y2="210" stroke="#3a3f46"/><line x1="40" y1="40" x2="40" y2="210" stroke="#3a3f46"/><line x1="40" y1="210.0" x2="720" y2="210.0" stroke="#23262b"/><text x="34" y="214.0" fill="#8b939d" font-size="11" text-anchor="end">2.32</text><line x1="40" y1="167.5" x2="720" y2="167.5" stroke="#23262b"/><text x="34" y="171.5" fill="#8b939d" font-size="11" text-anchor="end">2.35</text><line x1="40" y1="125.0" x2="720" y2="125.0" stroke="#23262b"/><text x="34" y="129.0" fill="#8b939d" font-size="11" text-anchor="end">2.37</text><line x1="40" y1="82.5" x2="720" y2="82.5" stroke="#23262b"/><text x="34" y="86.5" fill="#8b939d" font-size="11" text-anchor="end">2.40</text><line x1="40" y1="40.0" x2="720" y2="40.0" stroke="#23262b"/><text x="34" y="44.0" fill="#8b939d" font-size="11" text-anchor="end">2.43</text><text x="40.0" y="226" fill="#8b939d" font-size="11" text-anchor="middle">7150</text><text x="380.0" y="226" fill="#8b939d" font-size="11" text-anchor="middle">8615</text><text x="720.0" y="226" fill="#8b939d" font-size="11" text-anchor="middle">10080</text><text x="380" y="244" fill="#8b939d" font-size="11" text-anchor="middle">native step</text><path d="M42.3 101.3 L44.6 108.3 L47.0 95.7 L49.3 112.3 L51.6 99.8 L53.9 47.8 L56.2 79.8 L58.6 71.6 L60.9 113.0 L63.2 98.8 L65.5 112.2 L67.8 130.3 L70.2 112.1 L72.5 104.4 L74.8 82.1 L77.1 114.5 L79.5 116.5 L81.8 72.7 L84.1 79.3 L86.4 103.5 L88.7 79.8 L91.1 121.7 L93.4 75.3 L95.7 124.5 L98.0 131.7 L100.3 90.9 L102.7 100.2 L105.0 85.8 L107.3 86.8 L109.6 78.6 L111.9 99.5 L114.3 80.1 L116.6 121.8 L118.9 138.7 L121.2 92.3 L123.5 114.8 L125.9 103.2 L128.2 119.8 L130.5 100.9 L132.8 113.2 L135.2 111.3 L137.5 88.0 L139.8 109.2 L142.1 111.4 L144.4 103.9 L146.8 133.8 L149.1 114.1 L151.4 95.3 L153.7 111.0 L156.0 135.4 L158.4 117.9 L160.7 97.1 L163.0 112.2 L165.3 92.5 L167.6 116.7 L170.0 90.4 L172.3 115.1 L174.6 97.7 L176.9 97.9 L179.2 139.0 L181.6 111.4 L183.9 105.2 L186.2 128.0 L188.5 106.1 L190.9 117.2 L193.2 110.9 L195.5 113.9 L197.8 127.7 L200.1 125.1 L202.5 119.3 L204.8 121.1 L207.1 136.9 L209.4 111.8 L211.7 123.5 L214.1 97.6 L216.4 84.7 L218.7 134.5 L221.0 103.1 L223.3 113.8 L225.7 119.9 L228.0 121.9 L230.3 128.1 L232.6 145.4 L234.9 88.1 L237.3 121.1 L239.6 131.9 L241.9 121.1 L244.2 158.4 L246.6 115.0 L248.9 92.6 L251.2 156.7 L253.5 108.6 L255.8 140.1 L258.2 113.6 L260.5 104.4 L262.8 154.9 L265.1 126.0 L267.4 86.1 L269.8 120.5 L272.1 143.6 L274.4 126.8 L276.7 126.2 L279.0 129.8 L281.4 107.1 L283.7 139.2 L286.0 162.3 L288.3 131.4 L290.6 143.9 L293.0 135.3 L295.3 137.0 L297.6 124.8 L299.9 136.5 L302.3 128.3 L304.6 118.5 L306.9 92.7 L309.2 109.7 L311.5 141.5 L313.9 152.6 L316.2 120.4 L318.5 110.1 L320.8 115.9 L323.1 107.8 L325.5 155.7 L327.8 149.1 L330.1 130.3 L332.4 141.0 L334.7 177.6 L337.1 171.4 L339.4 139.7 L341.7 147.3 L344.0 152.8 L346.3 111.0 L348.7 147.7 L351.0 145.8 L353.3 151.5 L355.6 164.5 L358.0 122.9 L360.3 138.9 L362.6 159.2 L364.9 159.4 L367.2 139.4 L369.6 133.0 L371.9 137.8 L374.2 156.4 L376.5 154.2 L378.8 105.8 L381.2 120.4 L383.5 161.5 L385.8 143.9 L388.1 161.5 L390.4 160.8 L392.8 130.2 L395.1 152.1 L397.4 145.1 L399.7 131.7 L402.0 147.9 L404.4 157.9 L406.7 124.1 L409.0 138.2 L411.3 157.5 L413.7 136.8 L416.0 175.9 L418.3 137.0 L420.6 119.7 L422.9 170.6 L425.3 182.1 L427.6 133.1 L429.9 154.8 L432.2 131.7 L434.5 163.8 L436.9 156.6 L439.2 152.0 L441.5 147.9 L443.8 143.0 L446.1 169.6 L448.5 137.0 L450.8 143.5 L453.1 170.0 L455.4 176.3 L457.7 169.8 L460.1 151.8 L462.4 119.3 L464.7 155.2 L467.0 134.0 L469.4 168.9 L471.7 158.9 L474.0 163.0 L476.3 149.7 L478.6 134.5 L481.0 156.1 L483.3 139.1 L485.6 136.0 L487.9 158.5 L490.2 160.9 L492.6 154.9 L494.9 137.6 L497.2 158.1 L499.5 153.7 L501.8 139.4 L504.2 142.9 L506.5 144.0 L508.8 150.4 L511.1 165.3 L513.4 161.1 L515.8 143.7 L518.1 144.7 L520.4 162.1 L522.7 155.9 L525.1 145.1 L527.4 150.7 L529.7 164.4 L532.0 170.3 L534.3 168.4 L536.7 159.3 L539.0 142.9 L541.3 166.7 L543.6 165.6 L545.9 199.0 L548.3 160.4 L550.6 155.2 L552.9 167.3 L555.2 141.6 L557.5 161.4 L559.9 139.7 L562.2 182.3 L564.5 186.0 L566.8 179.1 L569.1 192.4 L571.5 166.6 L573.8 190.1 L576.1 170.2 L578.4 148.9 L580.8 179.6 L583.1 150.0 L585.4 160.3 L587.7 197.8 L590.0 168.9 L592.4 182.0 L594.7 172.6 L597.0 191.3 L599.3 177.3 L601.6 171.4 L604.0 159.0 L606.3 176.3 L608.6 154.1 L610.9 150.2 L613.2 174.7 L615.6 167.3 L617.9 161.0 L620.2 180.0 L622.5 166.5 L624.8 186.0 L627.2 184.7 L629.5 170.5 L631.8 163.2 L634.1 145.2 L636.5 159.6 L638.8 166.3 L641.1 188.7 L643.4 165.3 L645.7 155.4 L648.1 184.3 L650.4 177.1 L652.7 186.0 L655.0 179.9 L657.3 166.0 L659.7 167.2 L662.0 179.7 L664.3 160.9 L666.6 169.7 L668.9 163.7 L671.3 184.4 L673.6 187.9 L675.9 190.6 L678.2 160.0 L680.5 165.2 L682.9 131.6 L685.2 151.2 L687.5 177.7 L689.8 188.5 L692.2 168.2 L694.5 184.8 L696.8 191.2 L699.1 171.3 L701.4 194.6 L703.8 170.0 L706.1 202.2 L708.4 190.7 L710.7 199.0 L713.0 169.2 L715.4 178.2 L717.7 191.0 L720.0 181.6" fill="none" stroke="#7fb2ff" stroke-width="1.8"/><line x1="703.8" y1="40" x2="703.8" y2="210" stroke="#e8c37a" stroke-dasharray="4 4"/><text x="708.8" y="52" fill="#e8c37a" font-size="11">10010</text></svg></figure>
175
+ <p class="note">train/loss against native step, read from the <a href="https://track.fabryka.ai/run/f056bb5c-f1da-56d5-9776-cc33063f10a5">public continuation run</a> through its public API at 2026-10-28T00:35Z. Dashed line marks the 8×L40S adoption at milestone 10010. The curve is measured, not fitted.</p>
176
+ <figure><svg viewBox="0 0 760 250" role="img" style="width:100%;height:auto;background:transparent" xmlns="http://www.w3.org/2000/svg"><line x1="40" y1="210" x2="720" y2="210" stroke="#3a3f46"/><line x1="40" y1="40" x2="40" y2="210" stroke="#3a3f46"/><line x1="40" y1="210.0" x2="720" y2="210.0" stroke="#23262b"/><text x="34" y="214.0" fill="#8b939d" font-size="11" text-anchor="end">41</text><line x1="40" y1="167.5" x2="720" y2="167.5" stroke="#23262b"/><text x="34" y="171.5" fill="#8b939d" font-size="11" text-anchor="end">56</text><line x1="40" y1="125.0" x2="720" y2="125.0" stroke="#23262b"/><text x="34" y="129.0" fill="#8b939d" font-size="11" text-anchor="end">71</text><line x1="40" y1="82.5" x2="720" y2="82.5" stroke="#23262b"/><text x="34" y="86.5" fill="#8b939d" font-size="11" text-anchor="end">85</text><line x1="40" y1="40.0" x2="720" y2="40.0" stroke="#23262b"/><text x="34" y="44.0" fill="#8b939d" font-size="11" text-anchor="end">100</text><text x="40.0" y="226" fill="#8b939d" font-size="11" text-anchor="middle">7200</text><text x="380.0" y="226" fill="#8b939d" font-size="11" text-anchor="middle">8025</text><text x="720.0" y="226" fill="#8b939d" font-size="11" text-anchor="middle">8850</text><text x="380" y="244" fill="#8b939d" font-size="11" text-anchor="middle">native step</text><path d="M40.0 113.1 L40.0 113.1 L60.6 112.9 L60.6 112.9 L81.2 114.7 L81.2 114.7 L101.8 112.9 L101.8 112.9 L122.4 112.3 L143.0 113.6 L163.6 114.6 L184.2 111.5 L204.8 111.7 L225.5 111.4 L246.1 112.2 L266.7 114.3 L287.3 113.2 L307.9 112.9 L328.5 111.2 L349.1 111.9 L369.7 112.7 L390.3 114.5 L410.9 114.7 L431.5 113.2 L452.1 115.0 L472.7 114.7 L493.3 114.0 L513.9 112.5 L534.5 111.2 L555.2 113.8 L575.8 113.5 L596.4 113.4 L617.0 113.3 L637.6 111.5 L658.2 112.5 L678.8 112.5 L699.4 110.9 L720.0 111.5" fill="none" stroke="#8fe0ad" stroke-width="1.8"/><path d="M40.0 184.8 L40.0 184.8 L60.6 184.4 L60.6 184.4 L81.2 186.4 L81.2 186.4 L101.8 184.8 L101.8 184.8 L122.4 188.0 L143.0 184.2 L163.6 187.0 L184.2 186.5 L204.8 187.1 L225.5 187.0 L246.1 184.4 L266.7 188.0 L287.3 186.4 L307.9 188.8 L328.5 188.3 L349.1 183.8 L369.7 184.9 L390.3 187.4 L410.9 187.1 L431.5 184.0 L452.1 185.9 L472.7 184.9 L493.3 183.7 L513.9 184.9 L534.5 181.4 L555.2 184.7 L575.8 183.1 L596.4 185.4 L617.0 184.7 L637.6 185.2 L658.2 185.4 L678.8 183.3 L699.4 179.7 L720.0 183.7" fill="none" stroke="#7fb2ff" stroke-width="1.8"/><path d="M40.0 44.6 L60.6 44.4 L81.2 44.1 L101.8 44.1 L122.4 43.9 L143.0 44.2 L163.6 45.3 L184.2 44.1 L204.8 43.9 L225.5 43.5 L246.1 43.7 L266.7 44.1 L287.3 44.0 L307.9 43.4 L328.5 43.7 L349.1 43.3 L369.7 43.2 L390.3 43.5 L410.9 43.3 L431.5 43.5 L452.1 43.2 L472.7 43.1 L493.3 44.1 L513.9 43.6 L534.5 44.3 L555.2 43.9 L575.8 44.2 L596.4 44.1 L617.0 43.9 L637.6 44.1 L658.2 43.9 L678.8 43.9 L699.4 43.7 L720.0 44.8" fill="none" stroke="#e8c37a" stroke-width="1.8"/><path d="M40.0 203.9 L60.6 205.5 L81.2 203.8 L101.8 203.6 L122.4 205.4 L143.0 203.5 L163.6 204.4 L184.2 202.4 L204.8 203.1 L225.5 203.8 L246.1 203.6 L266.7 204.3 L287.3 201.0 L307.9 204.2 L328.5 204.4 L349.1 202.9 L369.7 204.1 L390.3 204.7 L410.9 206.4 L431.5 203.8 L452.1 203.5 L472.7 202.4 L493.3 202.0 L513.9 201.5 L534.5 204.3 L555.2 202.7 L575.8 201.5 L596.4 202.1 L617.0 202.0 L637.6 200.1 L658.2 204.4 L678.8 205.6 L699.4 202.2 L720.0 202.4" fill="none" stroke="#f2a0a0" stroke-width="1.8"/><rect x="48" y="46" width="10" height="10" fill="#8fe0ad"/><text x="62" y="55" fill="#c9d1d9" font-size="11">BLiMP EN</text><rect x="198" y="46" width="10" height="10" fill="#7fb2ff"/><text x="212" y="55" fill="#c9d1d9" font-size="11">ARC-Easy EN</text><rect x="348" y="46" width="10" height="10" fill="#e8c37a"/><text x="362" y="55" fill="#c9d1d9" font-size="11">MultiBLiMP PL</text><rect x="498" y="46" width="10" height="10" fill="#f2a0a0"/><text x="512" y="55" fill="#c9d1d9" font-size="11">ARC-Easy-PL</text></svg></figure>
177
+ <p class="note">Automatic mid-run evaluation, one instrument (lm-eval harness), every 50 steps: BLiMP EN, ARC-Easy EN, MultiBLiMP PL and ARC-Easy-PL in percent. Preliminary points from a run in progress — not final results, and not directly comparable with the offline table above (different procedure).</p>
178
+ </section>
179
+
180
  <section>
181
  <h2>How we do it</h2>
182
  <ol class="steps">
 
200
  <section>
201
  <h2>What's next</h2>
202
  <p><b>Corpus continuation, 1.04–1.06:</b> the global trajectory runs from native step 7150 toward the target 13848.
203
+ Checkpoint milestones are 10010 (1.04), 12870 (1.05) and 13848 (1.06). Milestone 10010 is verified and reached; training now runs on 8×L40S at ~99k tokens/s.
204
  <a href="https://track.fabryka.ai/run/f056bb5c-f1da-56d5-9776-cc33063f10a5">Follow the live continuation</a>.</p>
205
  <p class="note">The linked runs belong to the author's Track account. The completed 1.03 run includes its resumed steps;
206
  corpus continuation and benchmark history are separate records of the same model lineage.</p>
207
+ <p><b>Beyond 13848 — roadmap candidates, not scheduled runs:</b> a final learning-rate decay, and extending the context
208
+ window in stages (first 32k tokens, up to 256k as the target), each stage checked with long-context tests. The 1.04 plan ends at
209
+ 13848 on a constant learning rate; these follow-ups require their own plan, cost estimate and go-ahead, and a decay over an
210
+ exhausted corpus would need repeated or new data.</p>
211
  <p class="note">These are base models: they predict the continuation of a text, they were not tuned for chat or for
212
  following instructions, and they can be wrong.</p>
213
  </section>
 
264
  <a href="https://track.fabryka.ai/run/1426c0a8-2f07-5a03-80fc-92f7e922aa4b">benchmarki checkpointów</a> ·
265
  <a href="https://huggingface.co/SlayerLab/Slayer1B-1.03-checkpoints">archiwum checkpointów natywnych — aktualne pliki</a>
266
  </div>
267
+ <div class="card card-status">
268
+ <h3>Slayer1B 1.04 <span class="tag live">trening w toku</span></h3>
269
+ <div class="meta">kontynuacja korpusu 7150 → 10010 · cel ~10,5 mld tokenów</div>
270
+ <p>Start na 4×L40S, adopcja na 8×L40S przy kamieniu milowym 10010 po niezależnej weryfikacji SHA-256
271
+ checkpointu (trzy źródła) i bramce NCCL/tempo (~100 tys. tokenów/s na syntetyku).</p>
272
+ <p class="note">SHA-256 kamienia 10010: <code class="sha">04438959d7b317d1a8d972b264e9610a06fd8addde7349c6e9f458ee56dc1e51</code></p>
273
+ <a href="https://track.fabryka.ai/run/f056bb5c-f1da-56d5-9776-cc33063f10a5">bieg kontynuacji na żywo</a> ·
274
+ <a href="https://huggingface.co/SlayerLab/Slayer1B-1.04-checkpoints">archiwum checkpointów natywnych — aktualne pliki</a>
275
+ </div>
276
  </div>
277
  </section>
278
 
 
315
  <p class="note">Same pomiary, bez dopasowań: punkt kontrolny co 150 kroków. Historyczne dane: <a href="https://huggingface.co/spaces/SlayerLab/Slayer1B/resolve/b67cc3f1c3a56ea16f2ae92f44023e70750e371d/scores-history.csv">scores-history.csv</a>.</p>
316
  </section>
317
 
318
+ <section>
319
+ <h2>Ewaluacja w trakcie biegu: kontynuacja 1.04</h2>
320
+ <div class="wrap">
321
+ <table>
322
+ <tr><th>test</th><th>krok 7200</th><th>krok 7250</th></tr>
323
+ <tr><td>BLiMP (EN, gramatyka), %</td><td>74,63</td><td class="best">74,70</td></tr>
324
+ <tr><td>ARC-Easy (EN), %</td><td>49,75</td><td class="best">49,87</td></tr>
325
+ <tr><td>WikiText-2, perpleksja na bajt ↓</td><td class="best">2,3439</td><td>2,3456</td></tr>
326
+ <tr><td>WikiText-2, perpleksja na token ↓</td><td class="best">33,36</td><td>33,45</td></tr>
327
+ <tr><td>MultiBLiMP (PL, gramatyka), %</td><td>98,41</td><td class="best">98,47</td></tr>
328
+ <tr><td>ARC-Easy-PL, %</td><td class="best">43,10</td><td>42,55</td></tr>
329
+ </table>
330
+ </div>
331
+ <p class="note">Wstępne punkty w trakcie biegu z <a href="https://track.fabryka.ai/run/f056bb5c-f1da-56d5-9776-cc33063f10a5">kontynuacji 1.04 na żywo</a>,
332
+ mierzone automatycznie podczas treningu (harness lm-eval, jeden przyrząd). To nie są wyniki końcowe i nie są wprost
333
+ porównywalne z tabelą offline powyżej (inna procedura). Oba punkty pochodzą z jednego środowiska, więc 7200 → 7250
334
+ to czyste porównanie w obrębie biegu; przy 50 krokach różnicy zmiany są małe i mieszczą się w szumie biegu.</p>
335
+ </section>
336
+
337
+ <section>
338
+ <h2>Wykresy na żywo: kontynuacja 1.04</h2>
339
+ <figure><svg viewBox="0 0 760 250" role="img" style="width:100%;height:auto;background:transparent" xmlns="http://www.w3.org/2000/svg"><line x1="40" y1="210" x2="720" y2="210" stroke="#3a3f46"/><line x1="40" y1="40" x2="40" y2="210" stroke="#3a3f46"/><line x1="40" y1="210.0" x2="720" y2="210.0" stroke="#23262b"/><text x="34" y="214.0" fill="#8b939d" font-size="11" text-anchor="end">2.32</text><line x1="40" y1="167.5" x2="720" y2="167.5" stroke="#23262b"/><text x="34" y="171.5" fill="#8b939d" font-size="11" text-anchor="end">2.35</text><line x1="40" y1="125.0" x2="720" y2="125.0" stroke="#23262b"/><text x="34" y="129.0" fill="#8b939d" font-size="11" text-anchor="end">2.37</text><line x1="40" y1="82.5" x2="720" y2="82.5" stroke="#23262b"/><text x="34" y="86.5" fill="#8b939d" font-size="11" text-anchor="end">2.40</text><line x1="40" y1="40.0" x2="720" y2="40.0" stroke="#23262b"/><text x="34" y="44.0" fill="#8b939d" font-size="11" text-anchor="end">2.43</text><text x="40.0" y="226" fill="#8b939d" font-size="11" text-anchor="middle">7150</text><text x="380.0" y="226" fill="#8b939d" font-size="11" text-anchor="middle">8615</text><text x="720.0" y="226" fill="#8b939d" font-size="11" text-anchor="middle">10080</text><text x="380" y="244" fill="#8b939d" font-size="11" text-anchor="middle">krok native</text><path d="M42.3 101.3 L44.6 108.3 L47.0 95.7 L49.3 112.3 L51.6 99.8 L53.9 47.8 L56.2 79.8 L58.6 71.6 L60.9 113.0 L63.2 98.8 L65.5 112.2 L67.8 130.3 L70.2 112.1 L72.5 104.4 L74.8 82.1 L77.1 114.5 L79.5 116.5 L81.8 72.7 L84.1 79.3 L86.4 103.5 L88.7 79.8 L91.1 121.7 L93.4 75.3 L95.7 124.5 L98.0 131.7 L100.3 90.9 L102.7 100.2 L105.0 85.8 L107.3 86.8 L109.6 78.6 L111.9 99.5 L114.3 80.1 L116.6 121.8 L118.9 138.7 L121.2 92.3 L123.5 114.8 L125.9 103.2 L128.2 119.8 L130.5 100.9 L132.8 113.2 L135.2 111.3 L137.5 88.0 L139.8 109.2 L142.1 111.4 L144.4 103.9 L146.8 133.8 L149.1 114.1 L151.4 95.3 L153.7 111.0 L156.0 135.4 L158.4 117.9 L160.7 97.1 L163.0 112.2 L165.3 92.5 L167.6 116.7 L170.0 90.4 L172.3 115.1 L174.6 97.7 L176.9 97.9 L179.2 139.0 L181.6 111.4 L183.9 105.2 L186.2 128.0 L188.5 106.1 L190.9 117.2 L193.2 110.9 L195.5 113.9 L197.8 127.7 L200.1 125.1 L202.5 119.3 L204.8 121.1 L207.1 136.9 L209.4 111.8 L211.7 123.5 L214.1 97.6 L216.4 84.7 L218.7 134.5 L221.0 103.1 L223.3 113.8 L225.7 119.9 L228.0 121.9 L230.3 128.1 L232.6 145.4 L234.9 88.1 L237.3 121.1 L239.6 131.9 L241.9 121.1 L244.2 158.4 L246.6 115.0 L248.9 92.6 L251.2 156.7 L253.5 108.6 L255.8 140.1 L258.2 113.6 L260.5 104.4 L262.8 154.9 L265.1 126.0 L267.4 86.1 L269.8 120.5 L272.1 143.6 L274.4 126.8 L276.7 126.2 L279.0 129.8 L281.4 107.1 L283.7 139.2 L286.0 162.3 L288.3 131.4 L290.6 143.9 L293.0 135.3 L295.3 137.0 L297.6 124.8 L299.9 136.5 L302.3 128.3 L304.6 118.5 L306.9 92.7 L309.2 109.7 L311.5 141.5 L313.9 152.6 L316.2 120.4 L318.5 110.1 L320.8 115.9 L323.1 107.8 L325.5 155.7 L327.8 149.1 L330.1 130.3 L332.4 141.0 L334.7 177.6 L337.1 171.4 L339.4 139.7 L341.7 147.3 L344.0 152.8 L346.3 111.0 L348.7 147.7 L351.0 145.8 L353.3 151.5 L355.6 164.5 L358.0 122.9 L360.3 138.9 L362.6 159.2 L364.9 159.4 L367.2 139.4 L369.6 133.0 L371.9 137.8 L374.2 156.4 L376.5 154.2 L378.8 105.8 L381.2 120.4 L383.5 161.5 L385.8 143.9 L388.1 161.5 L390.4 160.8 L392.8 130.2 L395.1 152.1 L397.4 145.1 L399.7 131.7 L402.0 147.9 L404.4 157.9 L406.7 124.1 L409.0 138.2 L411.3 157.5 L413.7 136.8 L416.0 175.9 L418.3 137.0 L420.6 119.7 L422.9 170.6 L425.3 182.1 L427.6 133.1 L429.9 154.8 L432.2 131.7 L434.5 163.8 L436.9 156.6 L439.2 152.0 L441.5 147.9 L443.8 143.0 L446.1 169.6 L448.5 137.0 L450.8 143.5 L453.1 170.0 L455.4 176.3 L457.7 169.8 L460.1 151.8 L462.4 119.3 L464.7 155.2 L467.0 134.0 L469.4 168.9 L471.7 158.9 L474.0 163.0 L476.3 149.7 L478.6 134.5 L481.0 156.1 L483.3 139.1 L485.6 136.0 L487.9 158.5 L490.2 160.9 L492.6 154.9 L494.9 137.6 L497.2 158.1 L499.5 153.7 L501.8 139.4 L504.2 142.9 L506.5 144.0 L508.8 150.4 L511.1 165.3 L513.4 161.1 L515.8 143.7 L518.1 144.7 L520.4 162.1 L522.7 155.9 L525.1 145.1 L527.4 150.7 L529.7 164.4 L532.0 170.3 L534.3 168.4 L536.7 159.3 L539.0 142.9 L541.3 166.7 L543.6 165.6 L545.9 199.0 L548.3 160.4 L550.6 155.2 L552.9 167.3 L555.2 141.6 L557.5 161.4 L559.9 139.7 L562.2 182.3 L564.5 186.0 L566.8 179.1 L569.1 192.4 L571.5 166.6 L573.8 190.1 L576.1 170.2 L578.4 148.9 L580.8 179.6 L583.1 150.0 L585.4 160.3 L587.7 197.8 L590.0 168.9 L592.4 182.0 L594.7 172.6 L597.0 191.3 L599.3 177.3 L601.6 171.4 L604.0 159.0 L606.3 176.3 L608.6 154.1 L610.9 150.2 L613.2 174.7 L615.6 167.3 L617.9 161.0 L620.2 180.0 L622.5 166.5 L624.8 186.0 L627.2 184.7 L629.5 170.5 L631.8 163.2 L634.1 145.2 L636.5 159.6 L638.8 166.3 L641.1 188.7 L643.4 165.3 L645.7 155.4 L648.1 184.3 L650.4 177.1 L652.7 186.0 L655.0 179.9 L657.3 166.0 L659.7 167.2 L662.0 179.7 L664.3 160.9 L666.6 169.7 L668.9 163.7 L671.3 184.4 L673.6 187.9 L675.9 190.6 L678.2 160.0 L680.5 165.2 L682.9 131.6 L685.2 151.2 L687.5 177.7 L689.8 188.5 L692.2 168.2 L694.5 184.8 L696.8 191.2 L699.1 171.3 L701.4 194.6 L703.8 170.0 L706.1 202.2 L708.4 190.7 L710.7 199.0 L713.0 169.2 L715.4 178.2 L717.7 191.0 L720.0 181.6" fill="none" stroke="#7fb2ff" stroke-width="1.8"/><line x1="703.8" y1="40" x2="703.8" y2="210" stroke="#e8c37a" stroke-dasharray="4 4"/><text x="708.8" y="52" fill="#e8c37a" font-size="11">10010</text></svg></figure>
340
+ <p class="note">train/loss w funkcji kroku native, odczytane z <a href="https://track.fabryka.ai/run/f056bb5c-f1da-56d5-9776-cc33063f10a5">publicznego biegu kontynuacji</a> przez jego publiczne API o 2026-10-28T00:35Z. Linia przerywana oznacza adopcję na 8×L40S przy kamieniu 10010. Krzywa jest zmierzona, nie dopasowana.</p>
341
+ <figure><svg viewBox="0 0 760 250" role="img" style="width:100%;height:auto;background:transparent" xmlns="http://www.w3.org/2000/svg"><line x1="40" y1="210" x2="720" y2="210" stroke="#3a3f46"/><line x1="40" y1="40" x2="40" y2="210" stroke="#3a3f46"/><line x1="40" y1="210.0" x2="720" y2="210.0" stroke="#23262b"/><text x="34" y="214.0" fill="#8b939d" font-size="11" text-anchor="end">41</text><line x1="40" y1="167.5" x2="720" y2="167.5" stroke="#23262b"/><text x="34" y="171.5" fill="#8b939d" font-size="11" text-anchor="end">56</text><line x1="40" y1="125.0" x2="720" y2="125.0" stroke="#23262b"/><text x="34" y="129.0" fill="#8b939d" font-size="11" text-anchor="end">71</text><line x1="40" y1="82.5" x2="720" y2="82.5" stroke="#23262b"/><text x="34" y="86.5" fill="#8b939d" font-size="11" text-anchor="end">85</text><line x1="40" y1="40.0" x2="720" y2="40.0" stroke="#23262b"/><text x="34" y="44.0" fill="#8b939d" font-size="11" text-anchor="end">100</text><text x="40.0" y="226" fill="#8b939d" font-size="11" text-anchor="middle">7200</text><text x="380.0" y="226" fill="#8b939d" font-size="11" text-anchor="middle">8025</text><text x="720.0" y="226" fill="#8b939d" font-size="11" text-anchor="middle">8850</text><text x="380" y="244" fill="#8b939d" font-size="11" text-anchor="middle">krok native</text><path d="M40.0 113.1 L40.0 113.1 L60.6 112.9 L60.6 112.9 L81.2 114.7 L81.2 114.7 L101.8 112.9 L101.8 112.9 L122.4 112.3 L143.0 113.6 L163.6 114.6 L184.2 111.5 L204.8 111.7 L225.5 111.4 L246.1 112.2 L266.7 114.3 L287.3 113.2 L307.9 112.9 L328.5 111.2 L349.1 111.9 L369.7 112.7 L390.3 114.5 L410.9 114.7 L431.5 113.2 L452.1 115.0 L472.7 114.7 L493.3 114.0 L513.9 112.5 L534.5 111.2 L555.2 113.8 L575.8 113.5 L596.4 113.4 L617.0 113.3 L637.6 111.5 L658.2 112.5 L678.8 112.5 L699.4 110.9 L720.0 111.5" fill="none" stroke="#8fe0ad" stroke-width="1.8"/><path d="M40.0 184.8 L40.0 184.8 L60.6 184.4 L60.6 184.4 L81.2 186.4 L81.2 186.4 L101.8 184.8 L101.8 184.8 L122.4 188.0 L143.0 184.2 L163.6 187.0 L184.2 186.5 L204.8 187.1 L225.5 187.0 L246.1 184.4 L266.7 188.0 L287.3 186.4 L307.9 188.8 L328.5 188.3 L349.1 183.8 L369.7 184.9 L390.3 187.4 L410.9 187.1 L431.5 184.0 L452.1 185.9 L472.7 184.9 L493.3 183.7 L513.9 184.9 L534.5 181.4 L555.2 184.7 L575.8 183.1 L596.4 185.4 L617.0 184.7 L637.6 185.2 L658.2 185.4 L678.8 183.3 L699.4 179.7 L720.0 183.7" fill="none" stroke="#7fb2ff" stroke-width="1.8"/><path d="M40.0 44.6 L60.6 44.4 L81.2 44.1 L101.8 44.1 L122.4 43.9 L143.0 44.2 L163.6 45.3 L184.2 44.1 L204.8 43.9 L225.5 43.5 L246.1 43.7 L266.7 44.1 L287.3 44.0 L307.9 43.4 L328.5 43.7 L349.1 43.3 L369.7 43.2 L390.3 43.5 L410.9 43.3 L431.5 43.5 L452.1 43.2 L472.7 43.1 L493.3 44.1 L513.9 43.6 L534.5 44.3 L555.2 43.9 L575.8 44.2 L596.4 44.1 L617.0 43.9 L637.6 44.1 L658.2 43.9 L678.8 43.9 L699.4 43.7 L720.0 44.8" fill="none" stroke="#e8c37a" stroke-width="1.8"/><path d="M40.0 203.9 L60.6 205.5 L81.2 203.8 L101.8 203.6 L122.4 205.4 L143.0 203.5 L163.6 204.4 L184.2 202.4 L204.8 203.1 L225.5 203.8 L246.1 203.6 L266.7 204.3 L287.3 201.0 L307.9 204.2 L328.5 204.4 L349.1 202.9 L369.7 204.1 L390.3 204.7 L410.9 206.4 L431.5 203.8 L452.1 203.5 L472.7 202.4 L493.3 202.0 L513.9 201.5 L534.5 204.3 L555.2 202.7 L575.8 201.5 L596.4 202.1 L617.0 202.0 L637.6 200.1 L658.2 204.4 L678.8 205.6 L699.4 202.2 L720.0 202.4" fill="none" stroke="#f2a0a0" stroke-width="1.8"/><rect x="48" y="46" width="10" height="10" fill="#8fe0ad"/><text x="62" y="55" fill="#c9d1d9" font-size="11">BLiMP EN</text><rect x="198" y="46" width="10" height="10" fill="#7fb2ff"/><text x="212" y="55" fill="#c9d1d9" font-size="11">ARC-Easy EN</text><rect x="348" y="46" width="10" height="10" fill="#e8c37a"/><text x="362" y="55" fill="#c9d1d9" font-size="11">MultiBLiMP PL</text><rect x="498" y="46" width="10" height="10" fill="#f2a0a0"/><text x="512" y="55" fill="#c9d1d9" font-size="11">ARC-Easy-PL</text></svg></figure>
342
+ <p class="note">Automatyczna ewaluacja w trakcie biegu, jeden przyrząd (harness lm-eval), co 50 kroków: BLiMP EN, ARC-Easy EN, MultiBLiMP PL i ARC-Easy-PL w procentach. Wstępne punkty z biegu w toku — nie wyniki końcowe i nie porównywalne wprost z tabelą offline powyżej (inna procedura).</p>
343
+ </section>
344
+
345
  <section>
346
  <h2>Jak to robimy</h2>
347
  <ol class="steps">
 
365
  <section>
366
  <h2>Co dalej</h2>
367
  <p><b>Kontynuacja korpusu, 1.04–1.06:</b> globalna trajektoria prowadzi od kroku native 7150 do celu 13848.
368
+ Checkpointy wersji przypadają na 10010 (1.04), 12870 (1.05) i 13848 (1.06). Kamień 10010 jest zweryfikowany i osiągnięty; trening liczy teraz na 8×L40S przy ~99 tys. tokenów/s.
369
  <a href="https://track.fabryka.ai/run/f056bb5c-f1da-56d5-9776-cc33063f10a5">Postęp kontynuacji na żywo</a>.</p>
370
  <p class="note">Powiązane przebiegi są na koncie autora w Tracku. Zakończony run 1.03 obejmuje wznowione kroki;
371
  dalszy trening i historia benchmarków to osobne zapisy tej samej linii modelu.</p>
372
+ <p><b>Po 13848 — kandydaci roadmapy, nie zaplanowane biegi:</b> końcowe wygaszanie tempa uczenia oraz etapami wydłużanie okna
373
+ kontekstu (najpierw 32k tokenów, docelowo do 256k), każdy etap sprawdzany testami długiego kontekstu. Plan 1.04 kończy się na
374
+ 13848 przy stałym tempie uczenia; te kroki wymagają własnego planu, wyceny i zgody, a wygaszanie na wyczerpanym korpusie
375
+ wymagałoby powtórki części danych albo nowych danych.</p>
376
  <p class="note">To modele bazowe: przewidują dalszy ciąg tekstu, nie były dostrajane do rozmowy ani wykonywania poleceń
377
  i mogą się mylić.</p>
378
  </section>