Slayer1B 1.03: public model, owner's Track runs
Browse files- index.html +94 -32
index.html
CHANGED
|
@@ -32,6 +32,8 @@
|
|
| 32 |
tr:last-child td { border-bottom:none; }
|
| 33 |
td.best { color:var(--acc); font-weight:700; }
|
| 34 |
.note { color:var(--muted); font-size:.92rem; margin-top:10px; }
|
|
|
|
|
|
|
| 35 |
figure { margin:0; background:#fff; border-radius:14px; padding:8px; }
|
| 36 |
figure img { width:100%; display:block; }
|
| 37 |
ol.steps { padding-left:20px; } ol.steps li { margin:6px 0; }
|
|
@@ -51,15 +53,15 @@
|
|
| 51 |
<header>
|
| 52 |
<h1>Mission – Slayer 1B</h1>
|
| 53 |
<p class="by">Arkadiusz Słota – Fabryka AI (SlayerLab)</p>
|
| 54 |
-
<p class="lead">We are building a Polish language model of ~1 billion parameters from scratch and improving it
|
| 55 |
-
|
| 56 |
-
|
| 57 |
<div class="pills">
|
| 58 |
<span class="pill"><b>1,073,912,851</b> parameters</span>
|
| 59 |
<span class="pill">tokenizer <b>32,768</b>, Polish</span>
|
| 60 |
<span class="pill">context <b>4096</b> tokens</span>
|
| 61 |
<span class="pill">data: ~82% Polish, ~18% English</span>
|
| 62 |
-
<span class="pill">weights
|
| 63 |
</div>
|
| 64 |
</header>
|
| 65 |
|
|
@@ -87,6 +89,16 @@
|
|
| 87 |
<a href="https://huggingface.co/SlayerLab/Slayer1B-1.02-checkpoints">checkpoints</a> ·
|
| 88 |
<a href="https://track.fabryka.ai/run/21ae1234-82c2-42b7-887b-cbcc69901ae1">run</a>
|
| 89 |
</div>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 90 |
</div>
|
| 91 |
</section>
|
| 92 |
|
|
@@ -110,9 +122,23 @@
|
|
| 110 |
</section>
|
| 111 |
|
| 112 |
<section>
|
| 113 |
-
<h2>
|
| 114 |
-
<
|
| 115 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 116 |
</section>
|
| 117 |
|
| 118 |
<section>
|
|
@@ -120,13 +146,16 @@
|
|
| 120 |
<ol class="steps">
|
| 121 |
<li><b>From scratch.</b> A transformer (20 layers, width 2048), our own Polish tokenizer, Polish data with an English
|
| 122 |
minority (<a href="https://huggingface.co/datasets/SlayerLab/slayer-pl-8x3b">slayer-pl-8x3b</a>).</li>
|
| 123 |
-
<li><b>
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
<li><b>
|
| 128 |
-
|
| 129 |
-
|
|
|
|
|
|
|
|
|
|
| 130 |
</ol>
|
| 131 |
<p class="note">Technical report:
|
| 132 |
<a href="https://github.com/slayerlabs/Fabryka-AI-Notes/tree/main/papers/slayerlabs/2026-10-23_slayer1b-continuation">Slayer1B: Continuing a Polish-Dominant Base Model</a>.</p>
|
|
@@ -134,8 +163,11 @@
|
|
| 134 |
|
| 135 |
<section>
|
| 136 |
<h2>What's next</h2>
|
| 137 |
-
<p><b>
|
| 138 |
-
|
|
|
|
|
|
|
|
|
|
| 139 |
<p><b>At the end of the line:</b> one final learning-rate decay, then extending the context window in stages —
|
| 140 |
first to 32k tokens, with up to 256k as the target, each stage checked with long-context tests.</p>
|
| 141 |
<p class="note">These are base models: they predict the continuation of a text, they were not tuned for chat or for
|
|
@@ -148,15 +180,15 @@
|
|
| 148 |
<header>
|
| 149 |
<h1>Misja – Slayer 1B</h1>
|
| 150 |
<p class="by">Arkadiusz Słota – Fabryka AI (SlayerLab)</p>
|
| 151 |
-
<p class="lead">Budujemy od zera polski model językowy wielkości ~1 mld parametrów i ulepszamy go
|
| 152 |
-
|
| 153 |
-
|
| 154 |
<div class="pills">
|
| 155 |
<span class="pill"><b>1 073 912 851</b> parametrów</span>
|
| 156 |
<span class="pill">tokenizer <b>32 768</b>, polski</span>
|
| 157 |
<span class="pill">kontekst <b>4096</b> tokenów</span>
|
| 158 |
<span class="pill">dane: ~82 % PL, ~18 % EN</span>
|
| 159 |
-
<span class="pill">
|
| 160 |
</div>
|
| 161 |
</header>
|
| 162 |
|
|
@@ -184,6 +216,16 @@
|
|
| 184 |
<a href="https://huggingface.co/SlayerLab/Slayer1B-1.02-checkpoints">punkty kontrolne</a> ·
|
| 185 |
<a href="https://track.fabryka.ai/run/21ae1234-82c2-42b7-887b-cbcc69901ae1">przebieg</a>
|
| 186 |
</div>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 187 |
</div>
|
| 188 |
</section>
|
| 189 |
|
|
@@ -207,9 +249,23 @@
|
|
| 207 |
</section>
|
| 208 |
|
| 209 |
<section>
|
| 210 |
-
<h2>
|
| 211 |
-
<
|
| 212 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 213 |
</section>
|
| 214 |
|
| 215 |
<section>
|
|
@@ -217,13 +273,16 @@
|
|
| 217 |
<ol class="steps">
|
| 218 |
<li><b>Od zera.</b> Architektura typu transformer (20 warstw, szerokość 2048), własny polski tokenizer, polskie dane
|
| 219 |
z mniejszością angielską (<a href="https://huggingface.co/datasets/SlayerLab/slayer-pl-8x3b">slayer-pl-8x3b</a>).</li>
|
| 220 |
-
<li><b>
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
<li><b>
|
| 225 |
-
|
| 226 |
-
|
|
|
|
|
|
|
|
|
|
| 227 |
</ol>
|
| 228 |
<p class="note">Raport techniczny:
|
| 229 |
<a href="https://github.com/slayerlabs/Fabryka-AI-Notes/tree/main/papers/slayerlabs/2026-10-23_slayer1b-continuation">Slayer1B: Continuing a Polish-Dominant Base Model</a>.</p>
|
|
@@ -231,8 +290,11 @@
|
|
| 231 |
|
| 232 |
<section>
|
| 233 |
<h2>Co dalej</h2>
|
| 234 |
-
<p><b>
|
| 235 |
-
|
|
|
|
|
|
|
|
|
|
| 236 |
<p><b>Na końcu linii:</b> jedno końcowe wygaszanie tempa uczenia, a potem etapami wydłużanie okna kontekstu —
|
| 237 |
najpierw do 32k tokenów, docelowo do 256k, każdy etap sprawdzany testami długiego kontekstu.</p>
|
| 238 |
<p class="note">To modele bazowe: przewidują dalszy ciąg tekstu, nie były dostrajane do rozmowy ani wykonywania poleceń
|
|
|
|
| 32 |
tr:last-child td { border-bottom:none; }
|
| 33 |
td.best { color:var(--acc); font-weight:700; }
|
| 34 |
.note { color:var(--muted); font-size:.92rem; margin-top:10px; }
|
| 35 |
+
.sha { overflow-wrap:anywhere; font-size:.8rem; }
|
| 36 |
+
.card-status { grid-column:1 / -1; }
|
| 37 |
figure { margin:0; background:#fff; border-radius:14px; padding:8px; }
|
| 38 |
figure img { width:100%; display:block; }
|
| 39 |
ol.steps { padding-left:20px; } ol.steps li { margin:6px 0; }
|
|
|
|
| 53 |
<header>
|
| 54 |
<h1>Mission – Slayer 1B</h1>
|
| 55 |
<p class="by">Arkadiusz Słota – Fabryka AI (SlayerLab)</p>
|
| 56 |
+
<p class="lead">We are building a Polish language model of ~1 billion parameters from scratch and improving it through
|
| 57 |
+
checkpointed continuation. Released model weights are public; resumable checkpoint archives are published
|
| 58 |
+
separately. Follow each archive link for its current files. Training progress and measured evaluation results are reported separately below.</p>
|
| 59 |
<div class="pills">
|
| 60 |
<span class="pill"><b>1,073,912,851</b> parameters</span>
|
| 61 |
<span class="pill">tokenizer <b>32,768</b>, Polish</span>
|
| 62 |
<span class="pill">context <b>4096</b> tokens</span>
|
| 63 |
<span class="pill">data: ~82% Polish, ~18% English</span>
|
| 64 |
+
<span class="pill">released weights: <b>Apache-2.0</b></span>
|
| 65 |
</div>
|
| 66 |
</header>
|
| 67 |
|
|
|
|
| 89 |
<a href="https://huggingface.co/SlayerLab/Slayer1B-1.02-checkpoints">checkpoints</a> ·
|
| 90 |
<a href="https://track.fabryka.ai/run/21ae1234-82c2-42b7-887b-cbcc69901ae1">run</a>
|
| 91 |
</div>
|
| 92 |
+
<div class="card card-status">
|
| 93 |
+
<h3>Slayer1B 1.03 <span class="tag const">training complete</span></h3>
|
| 94 |
+
<div class="meta">native step 7150 · training loss 2.3955</div>
|
| 95 |
+
<p>Resume 6750 → 7150 completed on 4×L40S. The final 7150 model weights and lm-eval benchmark results are public.</p>
|
| 96 |
+
<p class="note">Checkpoint SHA-256: <code class="sha">d7cad9673f3ee706f3022ad6a97fa7593604806d4d25ba936c27885a8f115d97</code></p>
|
| 97 |
+
<a href="https://huggingface.co/SlayerLab/Slayer1B-1.03">model</a> ·
|
| 98 |
+
<a href="https://track.fabryka.ai/run/7ca4f84f-3b07-4525-8cde-f3a936dc3425">completed training run</a> ·
|
| 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 |
|
|
|
|
| 122 |
</section>
|
| 123 |
|
| 124 |
<section>
|
| 125 |
+
<h2>Additional benchmark history</h2>
|
| 126 |
+
<p><a href="https://track.fabryka.ai/run/1426c0a8-2f07-5a03-80fc-92f7e922aa4b">16 checkpoints from 1.00–1.03, through final 7150</a>
|
| 127 |
+
have separate lm-eval 0.4.13, zero-shot, fp32 results: 13 historical points plus 6900, 7050 and 7150.
|
| 128 |
+
These results use a different benchmark procedure from the table above.</p>
|
| 129 |
+
<p class="note">Accuracy and normalized accuracy are distinct series, in fractions (0–1); LAMBADA perplexity is separate.
|
| 130 |
+
The scan found n-gram overlap in shared/en for BoolQ, MMLU and ARC; its significance is still being assessed.
|
| 131 |
+
Overlap is not by itself a measured contamination rate. Open Charts and filter by <code>eval/</code>.
|
| 132 |
+
This is an imported result series, not a native automatic-benchmark job.</p>
|
| 133 |
+
<p class="note">The three new points use the same runner and task procedure, but a different runtime:
|
| 134 |
+
Torch 2.8.0+cu128, tokenizers 0.23.2 and NumPy 2.2.6, versus 2.14.1+cu126, 0.23.3 and 2.5.3 for the historical points.
|
| 135 |
+
The 6750 → 6900 boundary is not an identical-environment checkpoint comparison.</p>
|
| 136 |
+
</section>
|
| 137 |
+
|
| 138 |
+
<section>
|
| 139 |
+
<h2>Historical score curves: 1.00–1.02</h2>
|
| 140 |
+
<figure><img src="https://huggingface.co/spaces/SlayerLab/Slayer1B/resolve/b67cc3f1c3a56ea16f2ae92f44023e70750e371d/scores-history.png" alt="Evaluation history of Slayer1B 1.00, 1.01 and 1.02 versus training tokens"></figure>
|
| 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>
|
|
|
|
| 146 |
<ol class="steps">
|
| 147 |
<li><b>From scratch.</b> A transformer (20 layers, width 2048), our own Polish tokenizer, Polish data with an English
|
| 148 |
minority (<a href="https://huggingface.co/datasets/SlayerLab/slayer-pl-8x3b">slayer-pl-8x3b</a>).</li>
|
| 149 |
+
<li><b>Checkpointed continuation.</b> Earlier decayed versions branched from checkpoints before their final decay.
|
| 150 |
+
Version names mark milestones in one global data trajectory, not mandatory process restarts.
|
| 151 |
+
Retention maintenance resumed from verified checkpoint 7750 without resetting the data order.
|
| 152 |
+
The final learning-rate decay is planned once, at the end.</li>
|
| 153 |
+
<li><b>Source and model units.</b> The source corpus is counted as ~24 B <code>cl100k</code> tokens.
|
| 154 |
+
Our v6/32k tokenizer produces separate <code>uint16</code> train/validation files. Actual model-token counts and
|
| 155 |
+
stream boundaries come from pinned manifests. Next packs are prepared while the current stage trains.</li>
|
| 156 |
+
<li><b>We verify everything.</b> Checkpoint SHA-256 records are preserved outside the training machine.
|
| 157 |
+
Before local pruning, a full copy on a separate archive is verified and its receipt is persisted atomically.
|
| 158 |
+
From version 1.01 on, a loss forecast is written down before each run and evaluated afterwards.</li>
|
| 159 |
</ol>
|
| 160 |
<p class="note">Technical report:
|
| 161 |
<a href="https://github.com/slayerlabs/Fabryka-AI-Notes/tree/main/papers/slayerlabs/2026-10-23_slayer1b-continuation">Slayer1B: Continuing a Polish-Dominant Base Model</a>.</p>
|
|
|
|
| 163 |
|
| 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
|
|
|
|
| 180 |
<header>
|
| 181 |
<h1>Misja – Slayer 1B</h1>
|
| 182 |
<p class="by">Arkadiusz Słota – Fabryka AI (SlayerLab)</p>
|
| 183 |
+
<p class="lead">Budujemy od zera polski model językowy wielkości ~1 mld parametrów i ulepszamy go przez kontynuację
|
| 184 |
+
z punktów kontrolnych. Wagi wydanych modeli są publiczne; archiwa checkpointów do wznowienia publikujemy
|
| 185 |
+
osobno. Linki archiwów pokazują ich aktualne pliki. Postęp treningu i zmierzone wyniki ewaluacji pokazujemy poniżej oddzielnie.</p>
|
| 186 |
<div class="pills">
|
| 187 |
<span class="pill"><b>1 073 912 851</b> parametrów</span>
|
| 188 |
<span class="pill">tokenizer <b>32 768</b>, polski</span>
|
| 189 |
<span class="pill">kontekst <b>4096</b> tokenów</span>
|
| 190 |
<span class="pill">dane: ~82 % PL, ~18 % EN</span>
|
| 191 |
+
<span class="pill">opublikowane wagi: <b>Apache-2.0</b></span>
|
| 192 |
</div>
|
| 193 |
</header>
|
| 194 |
|
|
|
|
| 216 |
<a href="https://huggingface.co/SlayerLab/Slayer1B-1.02-checkpoints">punkty kontrolne</a> ·
|
| 217 |
<a href="https://track.fabryka.ai/run/21ae1234-82c2-42b7-887b-cbcc69901ae1">przebieg</a>
|
| 218 |
</div>
|
| 219 |
+
<div class="card card-status">
|
| 220 |
+
<h3>Slayer1B 1.03 <span class="tag const">trening zakończony</span></h3>
|
| 221 |
+
<div class="meta">krok native 7150 · strata treningowa 2,3955</div>
|
| 222 |
+
<p>Wznowienie 6750 → 7150 zakończone na 4×L40S. Wagi końcowego modelu 7150 i wyniki benchmarków lm-eval są publiczne.</p>
|
| 223 |
+
<p class="note">SHA-256 checkpointu: <code class="sha">d7cad9673f3ee706f3022ad6a97fa7593604806d4d25ba936c27885a8f115d97</code></p>
|
| 224 |
+
<a href="https://huggingface.co/SlayerLab/Slayer1B-1.03">model</a> ·
|
| 225 |
+
<a href="https://track.fabryka.ai/run/7ca4f84f-3b07-4525-8cde-f3a936dc3425">zakończony przebieg treningu</a> ·
|
| 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 |
|
|
|
|
| 249 |
</section>
|
| 250 |
|
| 251 |
<section>
|
| 252 |
+
<h2>Dodatkowa historia benchmarków</h2>
|
| 253 |
+
<p><a href="https://track.fabryka.ai/run/1426c0a8-2f07-5a03-80fc-92f7e922aa4b">16 checkpointów 1.00–1.03, do końcowego 7150</a>
|
| 254 |
+
ma osobne wyniki lm-eval 0.4.13, zero-shot, fp32: 13 historycznych punktów oraz 6900, 7050 i 7150.
|
| 255 |
+
To inna procedura benchmarkowa niż w tabeli powyżej.</p>
|
| 256 |
+
<p class="note">Trafność i trafność normalizowana to osobne serie, w ułamkach (0–1); perpleksja LAMBADA jest oddzielnie.
|
| 257 |
+
Skan wykrył nakładkę n-gramową w shared/en dla BoolQ, MMLU i ARC; ocena jej znaczenia nadal trwa.
|
| 258 |
+
Sama nakładka nie jest pomiarem odsetka rzeczywistego przecieku. W Charts wpisz w filtrze <code>eval/</code>.
|
| 259 |
+
To seria zaimportowanych wyników, nie natywny job automatycznej ewaluacji.</p>
|
| 260 |
+
<p class="note">Trzy nowe punkty mają ten sam runner i procedurę zadań, ale inne środowisko:
|
| 261 |
+
Torch 2.8.0+cu128, tokenizers 0.23.2 i NumPy 2.2.6 zamiast 2.14.1+cu126, 0.23.3 i 2.5.3 dla punktów historycznych.
|
| 262 |
+
Granica 6750 → 6900 nie jest porównaniem checkpointów w identycznym środowisku.</p>
|
| 263 |
+
</section>
|
| 264 |
+
|
| 265 |
+
<section>
|
| 266 |
+
<h2>Historyczne krzywe wyników: 1.00–1.02</h2>
|
| 267 |
+
<figure><img src="https://huggingface.co/spaces/SlayerLab/Slayer1B/resolve/b67cc3f1c3a56ea16f2ae92f44023e70750e371d/scores-history.png" alt="Historia wyników Slayer1B 1.00, 1.01 i 1.02 w funkcji liczby tokenów treningu"></figure>
|
| 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>
|
|
|
|
| 273 |
<ol class="steps">
|
| 274 |
<li><b>Od zera.</b> Architektura typu transformer (20 warstw, szerokość 2048), własny polski tokenizer, polskie dane
|
| 275 |
z mniejszością angielską (<a href="https://huggingface.co/datasets/SlayerLab/slayer-pl-8x3b">slayer-pl-8x3b</a>).</li>
|
| 276 |
+
<li><b>Kontynuacja z checkpointów.</b> Wcześniejsze wygaszane wersje odgałęziały się sprzed końcowego wygaszania.
|
| 277 |
+
Nazwy wersji oznaczają etapy jednej globalnej trajektorii danych, nie obowiązkowe restarty procesu.
|
| 278 |
+
Konserwacja retencji wznowiła trening ze zweryfikowanego checkpointu 7750 bez zerowania kolejności danych.
|
| 279 |
+
Końcowe wygaszanie tempa uczenia planujemy raz, na końcu.</li>
|
| 280 |
+
<li><b>Jednostki źródła i modelu.</b> Około 24 mld tokenów źródła liczymy przez <code>cl100k</code>.
|
| 281 |
+
Nasz tokenizer v6/32k tworzy osobne pliki train/val w <code>uint16</code>. Liczby tokenów modelu i granice
|
| 282 |
+
strumieni pochodzą z przypiętych manifestów. Następne wsady przygotowujemy podczas bieżącego treningu.</li>
|
| 283 |
+
<li><b>Wszystko sprawdzamy.</b> Sumy SHA-256 checkpointów zachowujemy poza maszyną treningową.
|
| 284 |
+
Przed lokalnym usunięciem pełna kopia na osobnym archiwum jest weryfikowana, a jej receipt utrwalany atomowo.
|
| 285 |
+
Od wersji 1.01 przed każdym biegiem zapisujemy prognozę straty i potem ją oceniamy.</li>
|
| 286 |
</ol>
|
| 287 |
<p class="note">Raport techniczny:
|
| 288 |
<a href="https://github.com/slayerlabs/Fabryka-AI-Notes/tree/main/papers/slayerlabs/2026-10-23_slayer1b-continuation">Slayer1B: Continuing a Polish-Dominant Base Model</a>.</p>
|
|
|
|
| 290 |
|
| 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
|