Misja - Slayer 1B: versions, results, evaluation history

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+scores-history.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,10 +1,16 @@
 ---
-title: Slayer1B
-emoji: 🐠
-colorFrom: red
-colorTo: indigo
+title: Misja – Slayer 1B
+emoji: 📈
+colorFrom: green
+colorTo: blue
 sdk: static
-pinned: false
+pinned: true
+license: apache-2.0
+short_description: Polish 1B LM trained from scratch, version by version
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Misja – Slayer 1B
+
+Arkadiusz Słota – Fabryka AI (SlayerLab). Polish ~1B-parameter base language model trained from scratch and continued
+version by version; versions, checkpoints, evaluation history and links. Numbers on the page come from the model cards
+and `scores-history.csv` (measured values only).

index.html

@@ -1,19 +1,139 @@
-<!doctype html>
-<html>
-	<head>
-		<meta charset="utf-8" />
-		<meta name="viewport" content="width=device-width" />
-		<title>My static Space</title>
-		<link rel="stylesheet" href="style.css" />
-	</head>
-	<body>
-		<div class="card">
-			<h1>Welcome to your static Space!</h1>
-			<p>You can modify this app directly by editing <i>index.html</i> in the Files and versions tab.</p>
-			<p>
-				Also don't forget to check the
-				<a href="https://huggingface.co/docs/hub/spaces" target="_blank">Spaces documentation</a>.
-			</p>
-		</div>
-	</body>
+<!DOCTYPE html>
+<html lang="pl">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>Misja – Slayer 1B</title>
+<style>
+  :root { --bg:#0f1115; --card:#171a21; --line:#2a2f3a; --fg:#e8eaee; --muted:#9aa3b2; --acc:#4fb477; --acc2:#e0565b; --acc3:#4c8bd6; }
+  * { box-sizing:border-box; }
+  body { margin:0; background:var(--bg); color:var(--fg); font:16px/1.6 system-ui,-apple-system,"Segoe UI",Roboto,sans-serif; }
+  main { max-width:980px; margin:0 auto; padding:40px 20px 64px; }
+  header h1 { font-size:2.4rem; margin:0 0 4px; letter-spacing:-.5px; }
+  header .by { color:var(--muted); margin:0 0 20px; }
+  .lead { font-size:1.12rem; max-width:760px; }
+  .pills { display:flex; flex-wrap:wrap; gap:8px; margin:18px 0 8px; }
+  .pill { background:var(--card); border:1px solid var(--line); border-radius:999px; padding:4px 12px; font-size:.9rem; color:var(--muted); }
+  .pill b { color:var(--fg); }
+  section { margin-top:44px; }
+  h2 { font-size:1.4rem; margin:0 0 12px; }
+  .cards { display:grid; grid-template-columns:repeat(auto-fit,minmax(260px,1fr)); gap:14px; }
+  .card { background:var(--card); border:1px solid var(--line); border-radius:14px; padding:18px; }
+  .card h3 { margin:0 0 2px; font-size:1.15rem; }
+  .card .meta { color:var(--muted); font-size:.92rem; margin-bottom:10px; }
+  .card a { color:var(--acc3); text-decoration:none; }
+  .card a:hover { text-decoration:underline; }
+  .tag { display:inline-block; font-size:.78rem; padding:2px 8px; border-radius:6px; margin-left:6px; vertical-align:middle; }
+  .tag.dec { background:#2b2140; color:#c7b2ff; } .tag.const { background:#1d3326; color:#8fe0ad; }
+  table { width:100%; border-collapse:collapse; background:var(--card); border:1px solid var(--line); border-radius:14px; overflow:hidden; }
+  th, td { padding:10px 12px; text-align:right; border-bottom:1px solid var(--line); font-variant-numeric:tabular-nums; }
+  th:first-child, td:first-child { text-align:left; }
+  th { color:var(--muted); font-weight:600; font-size:.9rem; }
+  tr:last-child td { border-bottom:none; }
+  td.best { color:var(--acc); font-weight:700; }
+  .note { color:var(--muted); font-size:.92rem; margin-top:10px; }
+  figure { margin:0; background:#fff; border-radius:14px; padding:8px; }
+  figure img { width:100%; display:block; }
+  ol.steps { padding-left:20px; } ol.steps li { margin:6px 0; }
+  .wrap { overflow-x:auto; }
+  footer { margin-top:56px; color:var(--muted); font-size:.9rem; border-top:1px solid var(--line); padding-top:16px; }
+  a { color:var(--acc3); }
+</style>
+</head>
+<body>
+<main>
+<header>
+  <h1>Misja – Slayer 1B</h1>
+  <p class="by">Arkadiusz Słota – Fabryka AI (SlayerLab)</p>
+  <p class="lead">Budujemy od zera polski model językowy wielkości ~1 mld parametrów i ulepszamy go wersja po wersji —
+  każdą wersję trenujemy dalej z tego miejsca, w którym skończyła poprzednia, na nowych danych. Wszystkie wersje,
+  punkty kontrolne i wyniki są publiczne.</p>
+  <div class="pills">
+    <span class="pill"><b>1 073 912 851</b> parametrów</span>
+    <span class="pill">tokenizer <b>32 768</b>, polski</span>
+    <span class="pill">kontekst <b>4096</b> tokenów</span>
+    <span class="pill">dane: ~82 % PL, ~18 % EN</span>
+    <span class="pill">licencja wag <b>Apache-2.0</b></span>
+  </div>
+</header>
+
+<section>
+  <h2>Wersje</h2>
+  <div class="cards">
+    <div class="card">
+      <h3>Slayer1B 1.00 <span class="tag dec">wygaszony</span></h3>
+      <div class="meta">1,5 mld tokenów treningu</div>
+      <a href="https://huggingface.co/SlayerLab/Slayer1B-1.00">model</a> ·
+      <a href="https://huggingface.co/SlayerLab/Slayer1B-1.00-checkpoints">punkty kontrolne</a> ·
+      <a href="https://track.fabryka.ai/run/48744ed2-3155-4294-b899-a6fda1f851da">przebieg</a>
+    </div>
+    <div class="card">
+      <h3>Slayer1B 1.01 <span class="tag dec">wygaszony</span></h3>
+      <div class="meta">3,0 mld tokenów · dalszy trening 1.00</div>
+      <a href="https://huggingface.co/SlayerLab/Slayer1B-1.01">model</a> ·
+      <a href="https://huggingface.co/SlayerLab/Slayer1B-1.01-checkpoints">punkty kontrolne</a> ·
+      <a href="https://track.fabryka.ai/run/4934429a-d393-42e0-8aac-03eae6b71ff7">przebieg</a>
+    </div>
+    <div class="card">
+      <h3>Slayer1B 1.02 <span class="tag const">stały LR</span></h3>
+      <div class="meta">6,0 mld tokenów · dalszy trening 1.01</div>
+      <a href="https://huggingface.co/SlayerLab/Slayer1B-1.02">model</a> ·
+      <a href="https://huggingface.co/SlayerLab/Slayer1B-1.02-checkpoints">punkty kontrolne</a> ·
+      <a href="https://track.fabryka.ai/run/21ae1234-82c2-42b7-887b-cbcc69901ae1">przebieg</a>
+    </div>
+  </div>
+</section>
+
+<section>
+  <h2>Wyniki</h2>
+  <div class="wrap">
+  <table>
+    <tr><th>test</th><th>1.00</th><th>1.01</th><th>1.02</th></tr>
+    <tr><td>tokeny treningu</td><td>1,5 mld</td><td>3,0 mld</td><td>6,0 mld</td></tr>
+    <tr><td>ARC-Easy (EN), %</td><td>42,89</td><td>47,98</td><td class="best">48,99</td></tr>
+    <tr><td>BLiMP (EN, gramatyka), %</td><td>71,51</td><td>73,63</td><td class="best">74,71</td></tr>
+    <tr><td>WikiText-2, perpleksja na bajt ↓</td><td>2,5857</td><td>2,4230</td><td class="best">2,3735</td></tr>
+    <tr><td>MultiBLiMP (PL, gramatyka), %</td><td>97,89</td><td>98,29</td><td class="best">98,38</td></tr>
+    <tr><td>ARC-Easy-PL, %</td><td>36,91</td><td>40,36</td><td class="best">41,84</td></tr>
+  </table>
+  </div>
+  <p class="note">Jeden przyrząd dla wszystkich wersji (protokół tablicy
+  <a href="https://track.fabryka.ai/models">track.fabryka.ai/models</a>, zero-shot). Porównanie 1.01 → 1.02 pytanie po pytaniu:
+  BLiMP +1,08 pp (95 % przedział [+0,78; +1,37]); na pozostałych testach kierunek dodatni, ale różnica nierozstrzygnięta.
+  1.00 i 1.01 kończyły się wygaszaniem tempa uczenia, 1.02 nie — więc 1.02 nie jest „czystym” porównaniem z 1.01.</p>
+</section>
+
+<section>
+  <h2>Jak rosną wyniki w trakcie treningu</h2>
+  <figure><img src="scores-history.png" alt="Historia wyników Slayer1B 1.00, 1.01 i 1.02 w funkcji liczby tokenów treningu"></figure>
+  <p class="note">Same pomiary, bez dopasowań: punkt kontrolny co 150 kroków. Dane: <a href="scores-history.csv">scores-history.csv</a>.</p>
+</section>
+
+<section>
+  <h2>Jak to robimy</h2>
+  <ol class="steps">
+    <li><b>Od zera.</b> Architektura typu transformer (20 warstw, szerokość 2048), własny polski tokenizer, polskie dane
+      z mniejszością angielską (<a href="https://huggingface.co/datasets/SlayerLab/slayer-pl-8x3b">slayer-pl-8x3b</a>).</li>
+    <li><b>Dalszy trening zamiast nowego.</b> Każda wersja startuje z punktu kontrolnego poprzedniej — sprzed końcowego
+      wygaszania tempa uczenia — i dostaje nową porcję danych, której model jeszcze nie widział.</li>
+    <li><b>Od 1.02 nie wygaszamy po drodze.</b> Trenujemy dalej na całym przygotowanym zasobie (~24 mld tokenów),
+      a końcowe wygaszanie zrobimy raz, na końcu.</li>
+    <li><b>Wszystko sprawdzamy.</b> Każdy punkt kontrolny ma sumę SHA-256 i jest pobierany z powrotem do kontroli;
+      przed każdym biegiem zapisujemy prognozę straty i potem uczciwie ją oceniamy.</li>
+  </ol>
+  <p class="note">Raport techniczny:
+  <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>
+</section>
+
+<section>
+  <h2>Co dalej</h2>
+  <p><b>Slayer1B 1.03</b> — kolejne 1,5 mld tokenów (razem 7,5 mld), dalszy trening 1.02 na naszych kartach 4×RTX 4090.
+  Postęp na żywo pojawi się na <a href="https://track.fabryka.ai/models">track.fabryka.ai</a>.</p>
+  <p class="note">To modele bazowe: przewidują dalszy ciąg tekstu, nie były dostrajane do rozmowy ani wykonywania poleceń
+  i mogą się mylić.</p>
+</section>
+
+<footer>Arkadiusz Słota – Fabryka AI (SlayerLab) · modele: <a href="https://huggingface.co/SlayerLab">huggingface.co/SlayerLab</a></footer>
+</main>
+</body>
 </html>

scores-history.csv

@@ -0,0 +1,50 @@
+model,step,tokens,checkpoint_sha256,arc_easy,blimp,wikitext2_token_ppl,wikitext2_byte_ppl,eff,multiblimp_pl,arc_easy_pl,eval_gpu
+1.00,150,157286400,8f113dd02835bf3cc6737980f1664464d24c26168c0b245e9b8ed365ce4fb6c9,28.37,53.56,821.2357,5.1032,54.63,67.45,27.19,A
+1.00,300,314572800,2765fc1b71989ac85cb44084e8083bd86cf6a33ba67b7df1453afdaabd355f16,31.61,58.92,165.6249,3.4591,59.81,86.09,28.75,A
+1.00,450,471859200,d940b909d2d0db48e4e0184405c96a472ec44f6d676798e358dd44a22c49f85f,35.69,62.87,106.3365,3.1062,63.13,91.99,30.22,B
+1.00,600,629145600,13b5e60a27e4732cb1f48d144d2651408d0a3bb4b6c6dabf5d165a664622df56,36.7,65.96,89.3185,2.9774,64.75,94.74,32.7,C
+1.00,750,786432000,3e5b03e6d546b139060503b932419c5ee04f1241822bb36b29951769f1a7e1d5,38.68,66.89,76.9765,2.8717,65.94,95.63,32.66,C
+1.00,900,943718400,6fa366b77f4d3460cd4ecaf5bf86f299eadf2bd3d38ac1728e7667095756099b,39.23,68.98,68.8196,2.7947,66.98,95.97,33.21,C
+1.00,1050,1101004800,236108e8f204995f4adcfea7d9a99437d230c1af12708c64c68c74afbc2910fd,39.44,70.13,63.0616,2.736,67.56,96.12,35.48,B
+1.00,1200,1258291200,2fe897999f7fc76a3805c9234e69f56df8743b03f13c0e0f12752d575fbc5102,41.08,70.19,57.951,2.6804,68.25,97.16,37.5,C
+1.00,1350,1415577600,551728562d7121f94980be587ed03dfb81619f968a91de1f29bbab9b42646cf7,42.05,71.28,51.9382,2.61,69.09,97.95,37.33,C
+1.00,1430,1499463680,7d092d4d0bf838da1115a5369bf0a6729fe4c170e0d5ad303873f7320d6c1866,42.89,71.51,49.9734,2.5857,69.5,97.89,36.91,C
+1.01,1200,1258291200,3637f1d90f60c2151724ad88b4b7c52925c87b7330a7680974b9093e8c1b4021,40.36,69.97,59.8334,2.7013,67.89,96.7,36.41,B
+1.01,1350,1415577600,d6d75584ffbd7261ed23f489ed4927d3e33fc3a6fef3bd9d3ad46b309b63daf4,40.87,70.3,56.5089,2.664,68.25,97.28,36.2,B
+1.01,1500,1572864000,3b2af30f8f4331f7bf549b1e7f0b2643ac2ec8ad3369a42b579d578930ba2934,41.2,70.72,54.4985,2.6407,68.55,97.34,36.49,B
+1.01,1650,1730150400,db254bfb77e302fc618f9cc7e3c4e23f1e3d6923386e54ad5fb930ff78a04bc4,44.44,70.79,52.6605,2.6188,69.7,97.52,37.29,B
+1.01,1800,1887436800,9591dfa0d41f68b9e5cc40db2caf6f601b102c31b37c76d2ddeffa9f26cec49e,42.59,71.46,49.8407,2.584,69.39,97.22,36.15,B
+1.01,1950,2044723200,6b8037e19b3fc42bb1c70599754749cdbb6a507a27fa4e2c3e316d676e969555,43.81,71.38,48.8777,2.5718,69.8,97.74,36.11,B
+1.01,2100,2202009600,b46f5e02f74dab2db1055fa009668c784411f403e01198dfff36deae9a0200f5,44.11,71.87,46.7259,2.5438,70.13,97.95,37.42,B
+1.01,2250,2359296000,aeddfb023abb11c8654b47c728c9680c232d1b6568697e3a1eee8128efd37054,45.75,71.45,45.6745,2.5298,70.57,97.65,38.26,B
+1.01,2400,2516582400,616df6070fbc6e338e106d4b535030d06767063168e80494b67f508005904b73,45.75,72.56,42.9297,2.492,71.03,98.04,38.93,B
+1.01,2550,2673868800,82fd3d07914308411a2ddec965ce6881f296b8bb6956369096534bdc2c45b0c8,45.5,73.37,40.2758,2.4537,71.31,98.29,38.59,B
+1.01,2700,2831155200,e2b5b29f05d3113c0bc95a5a4d74fe07fee057031068bc82844f129a341c09f0,46.93,73.67,39.4436,2.4413,71.91,98.44,40.07,B
+1.01,2850,2988441600,decd0f02487fa15c65aaaaf9e2c2320f196571efa0337c8c588705915f7fd315,47.6,73.61,38.2961,2.4238,72.16,98.26,40.45,B
+1.01,2860,2998927360,b0a288798ae614c38e37ff8176a293f6650cec2221d59e46d5116afac7c31eba,47.98,73.63,38.238,2.423,72.29,98.29,40.36,B
+1.02,2400,2516582400,289186300fa010c6918f4f6fd1fcf4e3054707f4d616200f52944d11d58109f8,44.7,72.08,44.9821,2.5205,70.45,97.77,39.18,D
+1.02,2400,2516582400,289186300fa010c6918f4f6fd1fcf4e3054707f4d616200f52944d11d58109f8,44.7,72.08,44.9821,2.5205,70.45,97.77,39.18,D
+1.02,2550,2673868800,14cd38933fcc2d6565fe9664d735cfc790c19aafb24d2d673f39dcca90a2ad6f,43.86,72.74,43.2078,2.4959,70.45,98.2,38.47,D
+1.02,2700,2831155200,7b6b7089f203825d413af43cbd06e8ef46f66e66bbf08267b60edd74dc48c25a,45.08,71.54,43.1767,2.4955,70.46,97.83,39.06,D
+1.02,2850,2988441600,c6883b8589a03f8859003f67cd65cf6ee82f185f3a52008380dcfa54b83c283e,45.33,72.11,42.0055,2.4789,70.77,97.95,39.14,D
+1.02,3000,3145728000,147fec8fca88dc3aa9e9dfd458cb5f95cba9b2352cafef77ad359d84168a7874,45.16,71.95,41.8042,2.476,70.67,98.17,39.94,D
+1.02,3150,3303014400,5f6f224c418953547de6fb65e12ecfbf3ca6fa803bb713396f4d315608e1e98e,45.33,72.69,41.9026,2.4774,70.97,98.04,40.78,D
+1.02,3300,3460300800,c8d6873290cffdf089cbdacc3700f5bda3ae5dd974d6442c13d430681057a225,45.88,73.69,40.6829,2.4597,71.52,98.23,39.69,D
+1.02,3450,3617587200,df68a26dd8f2b0b22d3acea4cb464d6fe1459a1ab2a2c3d720483e995a87b08c,44.99,72.82,39.9256,2.4485,70.97,98.59,39.94,D
+1.02,3600,3774873600,d22c2b63c1d8e676cf166637df477b6ccb9d2ba61de906d4da2418fa81834cbe,46.3,73.79,39.5912,2.4435,71.74,98.17,39.98,D
+1.02,3750,3932160000,b387d6d732b52a02a7b45b00a9853489467200996cfc335d7de6167234ba7d64,47.1,73.15,39.0532,2.4354,71.81,98.38,41.16,D
+1.02,3900,4089446400,1330992a4745a4ab72062ae052d3bdc867fd3682ecb1d181cc66b42b9d7cd8f4,46.93,73.89,38.71,2.4302,72.01,98.26,42.05,D
+1.02,4050,4246732800,84e9c06efe7fce214a9ffd4d676eb0ec777fc541838033503fc697b3a77f47a1,46.21,73.51,38.4269,2.4259,71.66,98.17,40.74,D
+1.02,4200,4404019200,f7517c553bae4c25131c0db8f3c3841ea218dcd8ad2eaf27d50ec5ec49a02925,47.31,73.7,37.9506,2.4185,72.11,98.56,40.87,D
+1.02,4350,4561305600,07d27aa81a056dd081493698cebaf4234759b40004555cad1bf07089374c9710,46.93,73.19,37.5648,2.4125,71.82,98.26,40.28,D
+1.02,4500,4718592000,642ec154019ea18bce4dae447e4eaad62bb34304e1be411657fa799148d3d84a,47.43,73.05,37.0296,2.4041,71.96,98.47,40.91,D
+1.02,4576,4798283776,423f51091c057336b3cfce80e6e64109c634c4efc742c503ca0588b7915ddff2,46.84,73.65,37.0717,2.4048,71.97,98.35,40.91,D
+1.02,4650,4875878400,7f3f03f5c63b9bfb655e16652e9685059fa766b14b3d161629a29e78f1cb35f0,47.01,73.35,37.1098,2.4054,71.92,98.17,40.74,D
+1.02,4800,5033164800,91f737dc34e91724fa9e4b25a1b1e743c4940dc000123baf77c0e3c2fd34334a,48.74,73.94,37.0214,2.404,72.7,98.41,40.4,D
+1.02,4950,5190451200,fa447bce4f93835752c2330fa251c38d12810452cf91e333a3dd96971addc77a,47.85,74.25,36.4382,2.3948,72.53,98.41,41.5,D
+1.02,5100,5347737600,a519d20a0cc409c715f2a4ebdca841f6f3bc0f3b0ee663600d00ee466b043e51,48.74,73.86,35.9729,2.3873,72.71,98.17,40.95,D
+1.02,5250,5505024000,f6faafb27bfbbcfa1122cc904edc4b3dd4a9e890702722af2f2b2eae7163d923,48.36,74.18,35.6959,2.3828,72.7,98.29,41.04,D
+1.02,5400,5662310400,7b2f6b18d9b1137c55db98a18f97b46b42f7ef0eb9d72eaeb513e72505e45626,48.32,73.67,35.6028,2.3813,72.52,98.66,41.62,D
+1.02,5550,5819596800,896d27baf2f60ebd46710aa420358500355d3d8b74281c1a689f698fbf189ade,49.41,74.38,34.8006,2.3682,73.16,98.26,41.54,D
+1.02,5700,5976883200,5a830fb09e3374caac8482198fb5ee7c1a3bea7fe8fd7d329cae5df38faa300d,48.99,74.61,34.7402,2.3672,73.1,98.41,41.84,D
+1.02,5720,5997854720,6b2a9b11df4c1063476d5235475c98d35abbab3be4282a83d9f3613bbcbc805a,48.99,74.71,35.1257,2.3735,73.11,98.38,41.84,D