File size: 25,784 Bytes
709c9a2 01d0891 709c9a2 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 | <!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta name="description" content="An interactive field guide to training Qwen, BERT and DeBERTa for BANKING77 on Apple Silicon.">
<title>Open Model Training Lab — Three Model Flight Recorder</title>
<link rel="stylesheet" href="styles.css">
</head>
<body>
<a class="skip-link" href="#main">Skip to the lesson</a>
<header class="topbar">
<a class="brand" href="#top" aria-label="Open Model Training Lab home">
<span class="brand-mark" aria-hidden="true">OM</span>
<span><strong>Training Lab</strong><small>Flight recorder · M2 Max</small></span>
</a>
<nav aria-label="Lesson navigation">
<a href="#map">Map</a>
<a href="#eras">Models</a>
<a href="#timeline">Journey</a>
<a href="#failures">Failures</a>
<a href="#interview">Interview</a>
</nav>
<button class="utility-button" id="depthToggle" type="button" aria-pressed="false">
<span class="utility-dot"></span> Show deeper notes
</button>
</header>
<main id="main">
<section class="hero" id="top" data-track="Welcome">
<div class="hero-copy">
<p class="eyebrow">Your project, translated into plain English</p>
<h1>You did not just <em>train a model.</em><br>You ran a controlled investigation.</h1>
<p class="hero-lede">We began with a 1.7-billion-parameter language model, discovered the limits of generative classification, moved to BERT and then DeBERTa, and learned that model architecture, numerical precision, data quality and honest evaluation all matter.</p>
<div class="hero-actions">
<a class="primary-action" href="#map">Start from the beginning <span aria-hidden="true">↓</span></a>
<button class="text-action" id="resumeButton" type="button">Resume where I stopped</button>
</div>
<dl class="hardware-strip" aria-label="Lab setup">
<div><dt>Machine</dt><dd>M2 Max · 32 GB</dd></div>
<div><dt>Frameworks</dt><dd>MLX + PyTorch MPS</dd></div>
<div><dt>Models</dt><dd>Qwen → BERT → DeBERTa</dd></div>
<div><dt>Task</dt><dd>77 banking intents</dd></div>
</dl>
</div>
<div class="recorder" aria-label="Project signal from raw request to canonical label">
<div class="recorder-head">
<span>LIVE CONCEPT TRACE</span>
<span class="recording"><i></i> REC 102</span>
</div>
<div class="signal-stage active">
<span class="signal-index">A</span>
<div><small>Human request</small><p>“The cash machine charged me an extra fee.”</p></div>
</div>
<div class="signal-wire"><span></span></div>
<div class="signal-stage">
<span class="signal-index">B</span>
<div><small>Architecture decision</small><p>Qwen generation → encoder classification</p></div>
</div>
<div class="signal-wire"><span></span></div>
<div class="signal-stage result">
<span class="signal-index">C</span>
<div><small>DeBERTa classifier head</small><code>cash_withdrawal_charge</code></div>
</div>
<p class="recorder-note">All three models already understood language. We changed how they were adapted to the same 77-intent decision problem.</p>
</div>
</section>
<section class="section map-section" id="map" data-track="Project map">
<div class="section-heading">
<p class="eyebrow">01 · First principles</p>
<h2>What are we actually building?</h2>
<p>A specialist classifier—not a chatbot, not RAG, and not a model trained from zero.</p>
</div>
<div class="concept-grid">
<article class="concept-card span-two">
<div class="card-label">The job</div>
<h3>Turn messy human language into one controlled label</h3>
<div class="transformation" aria-label="Classification example">
<div><span>INPUT</span><p>“My transfer is still waiting.”</p></div>
<svg viewBox="0 0 90 24" role="img" aria-label="becomes"><path d="M2 12h78M70 4l10 8-10 8"/></svg>
<div><span>OUTPUT</span><code>pending_transfer</code></div>
</div>
<p class="deep-note">Qwen solved this by generating label tokens. BERT and DeBERTa solve it with 77 output scores called logits. The highest score becomes the predicted intent.</p>
</article>
<article class="concept-card">
<div class="card-icon data-icon" aria-hidden="true"><i></i><i></i><i></i></div>
<div class="card-label">The curriculum</div>
<h3>BANKING77</h3>
<p>10,003 public training questions and 3,080 test questions across 77 banking intents.</p>
<button class="inline-reveal" type="button" data-reveal="datasetReveal">Why public data?</button>
<p class="reveal-copy" id="datasetReveal" hidden>It is reproducible, inspectable, and much cheaper than generating thousands of examples in chat. Anyone can rerun our experiment.</p>
</article>
<article class="concept-card">
<div class="card-icon model-icon" aria-hidden="true"><span>1.7B</span></div>
<div class="card-label">The first student</div>
<h3>Qwen3 1.7B</h3>
<p>A generative language model adapted with LoRA. It taught us the full training lifecycle, but generation was not the best architecture for closed 77-way classification.</p>
<p class="deep-note">“1.7B” means roughly 1.7 billion learned parameters. BF16 stores each weight with 16 bits. Qwen generated text, so invalid label strings were possible.</p>
</article>
<article class="concept-card">
<div class="card-icon adapter-icon" aria-hidden="true"><span></span><span></span></div>
<div class="card-label">What changed later</div>
<h3>Classifier fine-tuning</h3>
<p>BERT and DeBERTa use a dedicated 77-class head. We first trained the full encoder, then refined only selected upper layers.</p>
<p class="deep-note">A classifier head is a small final layer that converts the encoder's representation into 77 scores. Unlike generation, it can never produce a label outside those 77 positions.</p>
</article>
</div>
</section>
<section class="section era-section" id="eras" data-track="Three model eras">
<div class="section-heading compact">
<p class="eyebrow">02 · The architecture lesson</p>
<h2>Three models. One task. Three different ways to learn.</h2>
<p>Follow the lineage from a general-purpose generator to a purpose-built encoder classifier. The switch was an engineering decision, not an admission that the first model was useless.</p>
</div>
<div class="lineage-rail" aria-label="Model evolution from Qwen to BERT to DeBERTa">
<article class="era-card qwen-era">
<header><span>ERA A</span><strong>Qwen3 1.7B</strong></header>
<p class="era-role">Generative language model</p>
<div class="era-flow"><code>question</code><b>→</b><code>label tokens</code></div>
<dl><div><dt>Training</dt><dd>LoRA adapters</dd></div><div><dt>Best result</dt><dd>51.10% test</dd></div><div><dt>Main lesson</dt><dd>Complete LLM workflow</dd></div></dl>
<p class="deep-note">Qwen predicts the next token repeatedly. We needed constrained decoding because a reasonable phrase could still be an invalid BANKING77 label.</p>
</article>
<div class="lineage-junction"><span>Why switch?</span><p>A closed classification task does not need open-ended generation.</p></div>
<article class="era-card bert-era">
<header><span>ERA B</span><strong>BERT-Large</strong></header>
<p class="era-role">Bidirectional text encoder</p>
<div class="era-flow"><code>question</code><b>→</b><code>77 logits</code></div>
<dl><div><dt>Training</dt><dd>Full fine-tuning</dd></div><div><dt>Best result</dt><dd>91.56% validation</dd></div><div><dt>Main lesson</dt><dd>Architecture fit</dd></div></dl>
<p class="deep-note">BERT reads the entire question in both directions and compresses it into a representation. A new classifier head maps that representation directly to 77 choices.</p>
</article>
<div class="lineage-junction"><span>Why switch?</span><p>DeBERTa improves how content and token position are represented.</p></div>
<article class="era-card deberta-era current">
<header><span>ERA C · CHAMPION</span><strong>DeBERTa-v3-large</strong></header>
<p class="era-role">Stronger encoder classifier</p>
<div class="era-flow"><code>question</code><b>→</b><code>77 logits</code></div>
<dl><div><dt>Training</dt><dd>Full + upper-layer refinement</dd></div><div><dt>Champion</dt><dd>94.12% test</dd></div><div><dt>Latest</dt><dd>Exp015 rejected</dd></div></dl>
<p class="deep-note">DeBERTa separates token content from token position inside attention. We had to force float32 on MPS because lower-precision parameter updates became non-finite.</p>
</article>
</div>
<aside class="architecture-verdict">
<strong>The interview answer</strong>
<p>“I began with a generative LLM to learn SFT, LoRA and constrained decoding. Once the task was proven to be fixed-label classification, I moved to encoder models whose output layer directly represents the 77 intents. That architecture change improved accuracy by more than forty percentage points.”</p>
</aside>
</section>
<section class="section workbench-section" id="workbench" data-track="Model workbench">
<div class="section-heading compact">
<p class="eyebrow">03 · Model workbench</p>
<h2>Compare the checkpoints without mixing their meaning</h2>
<p>Pick a model state. Always read whether its score came from validation or reporting-only test data.</p>
</div>
<div class="workbench">
<div class="model-selector" role="tablist" aria-label="Model versions" id="modelSelector"></div>
<div class="instrument-panel">
<div class="instrument-copy">
<p class="instrument-kicker" id="modelKicker"></p>
<h3 id="modelName"></h3>
<p class="instrument-scope" id="modelScope"></p>
<p id="modelStory"></p>
<div class="decision-chip" id="modelDecision"></div>
</div>
<div class="gauges">
<div class="gauge-block">
<div class="gauge-label"><span>Recorded accuracy</span><strong id="accuracyValue"></strong></div>
<div class="gauge-track"><i id="accuracyBar"></i><b class="baseline-mark" title="Untouched baseline"></b></div>
<small id="accuracyNote"></small>
</div>
<div class="gauge-block">
<div class="gauge-label"><span>Macro F1</span><strong id="f1Value"></strong></div>
<div class="gauge-track secondary"><i id="f1Bar"></i></div>
<small>Every intent gets equal importance.</small>
</div>
<div class="gauge-block">
<div class="gauge-label"><span>Invalid outputs</span><strong id="invalidValue"></strong></div>
<div class="gauge-track danger"><i id="invalidBar"></i></div>
<small id="invalidNote"></small>
</div>
</div>
<dl class="model-facts" id="modelFacts"></dl>
</div>
</div>
<aside class="lesson-callout">
<span class="callout-number">!</span>
<div><strong>A score without its split is incomplete.</strong><p>We select encoder experiments using 770 validation examples. Test results are reporting evidence only and are now locked out of every future model decision.</p></div>
</aside>
</section>
<section class="section data-section" id="data" data-track="Data splits">
<div class="section-heading">
<p class="eyebrow">04 · Data discipline</p>
<h2>Three piles. Three different jobs.</h2>
<p>Most beginner ML mistakes come from letting information leak between these piles.</p>
</div>
<div class="split-lab">
<article class="split-card train">
<header><span>TRAIN</span><strong>9,233</strong></header>
<h3>The practice questions</h3>
<p>The optimizer sees these examples and updates either LoRA adapters or selected encoder weights.</p>
<div class="split-meter"><i style="--fill: 100%"></i></div>
</article>
<article class="split-card validate">
<header><span>VALIDATE</span><strong>770</strong></header>
<h3>The progress check</h3>
<p>Ten held-out examples per intent tell us whether training is learning or becoming unstable.</p>
<div class="split-meter"><i style="--fill: 32%"></i></div>
</article>
<article class="split-card test">
<header><span>TEST</span><strong>3,080</strong></header>
<h3>The final exam</h3>
<p>Never used to update weights. It was used for reporting on finalized champions and is now locked from future selection.</p>
<div class="split-meter"><i style="--fill: 58%"></i></div>
</article>
</div>
<div class="leakage-demo">
<div class="leakage-visual" aria-hidden="true"><span>TRAIN</span><i></i><span>TEST</span></div>
<div><h3>What is data leakage?</h3><p>It is like seeing exam answers while studying, then changing your study plan because of them. We verified zero text overlap, pinned checksums, and require every current experiment to print <code>test_rows_loaded: 0</code>.</p></div>
</div>
</section>
<section class="section timeline-section" id="timeline" data-track="Experiment journey">
<div class="section-heading compact">
<p class="eyebrow">05 · The flight recorder</p>
<h2>How we got here, decision by decision</h2>
<p>Select a stage to see what we did, why it mattered, and what it taught us.</p>
</div>
<div class="timeline-layout">
<ol class="timeline-nav" id="timelineNav" aria-label="Experiment stages"></ol>
<article class="timeline-detail" id="timelineDetail" aria-live="polite"></article>
</div>
</section>
<section class="section training-section" id="training" data-track="How training works">
<div class="section-heading">
<p class="eyebrow">06 · Inside one training step</p>
<h2>The same learning loop, with different moving weights</h2>
</div>
<div class="loop-diagram">
<article><span>1</span><h3>Encode</h3><p>Turn a banking question into token IDs and attach its correct class number.</p></article>
<div class="loop-arrow" aria-hidden="true">→</div>
<article><span>2</span><h3>Predict</h3><p>The model produces token probabilities or 77 class logits, depending on architecture.</p></article>
<div class="loop-arrow" aria-hidden="true">→</div>
<article><span>3</span><h3>Measure loss</h3><p>Cross-entropy measures how much probability the model failed to give the correct answer.</p></article>
<div class="loop-arrow" aria-hidden="true">→</div>
<article><span>4</span><h3>Update weights</h3><p>AdamW nudges adapters, the full encoder, or only unfrozen upper layers.</p></article>
<div class="loop-return" aria-hidden="true">Repeat with the next example ↺</div>
</div>
<div class="parameter-microscope">
<div>
<p class="card-label">LoRA microscope</p>
<h3>Frozen model, tiny moving parts</h3>
<p>Use the switch to compare the two adapter designs we tested.</p>
<div class="segmented" id="loraToggle" role="group" aria-label="LoRA target selection">
<button type="button" data-lora="qv" class="active">q + v</button>
<button type="button" data-lora="qkvo">q + k + v + o</button>
</div>
</div>
<div class="model-stack" id="modelStack" aria-label="Model parameter visualization"></div>
<dl class="microscope-stats">
<div><dt>Trainable values</dt><dd id="loraValues">917,504</dd></div>
<div><dt>Share of model</dt><dd id="loraShare">0.053%</dd></div>
<div><dt>Adapter file</dt><dd id="loraSize">3.68 MB</dd></div>
</dl>
</div>
</section>
<section class="section failure-section" id="failures" data-track="Failure lab">
<div class="section-heading compact">
<p class="eyebrow">07 · Failure lab</p>
<h2>The errors were part of the curriculum</h2>
<p>Open each incident. Learn to separate a code bug, a data problem, numerical instability, and an infrastructure limitation.</p>
</div>
<div class="failure-toolbar">
<div class="filter-group" id="failureFilters" role="group" aria-label="Filter failures"></div>
<span id="failureCount"></span>
</div>
<div class="failure-list" id="failureList"></div>
</section>
<section class="section metrics-section" id="metrics" data-track="Metrics">
<div class="section-heading">
<p class="eyebrow">08 · Reading the instruments</p>
<h2>Loss is not accuracy. Accuracy is not the whole story.</h2>
</div>
<div class="metric-cards">
<article><span class="metric-symbol">L</span><h3>Training loss</h3><p>How surprising the correct output tokens were during training. Useful for stability and convergence—not a final product score.</p></article>
<article><span class="metric-symbol">%</span><h3>Accuracy</h3><p>The fraction of test questions with the exact correct label. Easy to explain, but frequent classes can dominate.</p></article>
<article><span class="metric-symbol">F1</span><h3>Macro F1</h3><p>Compute F1 separately for each intent, then average. Every one of the 77 intents gets equal weight.</p></article>
<article><span class="metric-symbol">Ø</span><h3>Invalid rate</h3><p>Relevant to Qwen generation. Encoder classifier heads always choose one of their 77 output positions.</p></article>
</div>
<div class="metric-scenario">
<div class="scenario-copy"><p class="card-label">Think before revealing</p><h3>A model’s loss falls, but test accuracy falls too. Did training “work”?</h3></div>
<button class="primary-action small" type="button" data-reveal="metricAnswer">Reveal answer</button>
<p class="reveal-copy answer" id="metricAnswer" hidden>Optimization worked—the model learned its training objective better. Generalization did not. It may have overfit, learned noisy labels, or optimized an objective that does not improve held-out decisions. This is why validation selects experiments and the test set remains outside current development.</p>
</div>
</section>
<section class="section constraint-section" id="constraint" data-track="Constrained decoding">
<div class="section-heading compact">
<p class="eyebrow">09 · What Qwen taught us</p>
<h2>Constrained decoding puts guardrails on generation</h2>
</div>
<div class="constraint-demo">
<div class="token-tree">
<p class="card-label">Allowed token paths</p>
<div class="tree-root">start</div>
<div class="tree-branches">
<div><span>card_</span><i></i><small>arrival · linking · swallowed…</small></div>
<div><span>cash_</span><i></i><small>withdrawal_charge · …</small></div>
<div><span>pending_</span><i></i><small>transfer · top_up · …</small></div>
</div>
</div>
<div class="constraint-copy">
<p>At every generated token, we mask paths that cannot finish as one of the 77 labels.</p>
<div class="before-after">
<div><span>BEFORE</span><code>get_virtual_card</code><small>Meaningful, but invalid</small></div>
<div><span>AFTER</span><code>getting_virtual_card</code><small>Canonical taxonomy label</small></div>
</div>
<p class="deep-note">The mask changes logits before greedy selection. It never looks at the expected answer. Our matched batch-size experiment exists because changing batch shape can introduce tiny floating-point differences, so a fair A/B comparison must hold batch size constant.</p>
</div>
</div>
</section>
<section class="section noise-section" id="noise" data-track="Data quality investigation">
<div class="section-heading compact">
<p class="eyebrow">10 · Current investigation</p>
<h2>When “hard examples” may actually be noisy labels</h2>
<p>Exp012–015 changed our question from “How can we push the model harder?” to “Should the model trust every training label equally?”</p>
</div>
<div class="noise-pipeline">
<article><span>EXP012</span><strong>Audit</strong><p>Five-fold out-of-fold predictions examined 9,233 training rows without reading validation or test data.</p><b>1,186 disagreements</b></article>
<i aria-hidden="true">→</i>
<article class="rejected"><span>EXP013 / 014</span><strong>Hard negatives</strong><p>We retained original labels and pushed against likely rival labels. Both experiments lost one validation answer.</p><b>Hypothesis rejected</b></article>
<i aria-hidden="true">→</i>
<article class="rejected"><span>EXP015 · REJECTED</span><strong>Noise pruning</strong><p>Removed 1,078 confidently suspicious training rows and rewrote no labels. The stable run reached only 90.78% validation.</p><b>17 fewer correct than Exp011</b></article>
</div>
<div class="noise-rules">
<div><span>REMOVE</span><p>The OOF model disagrees and gives the supplied label less than 25% probability.</p></div>
<div><span>DO NOT RELABEL</span><p>A prediction is evidence of suspicion, not proof of the correct replacement label.</p></div>
<div><span>DO NOT TOUCH TEST</span><p>All preparation and probing report zero validation/test rows loaded.</p></div>
</div>
<p class="deep-note">Out-of-fold means each training example is scored by a model that did not train on that example. This reduces self-confirmation, but it does not make the predictions ground truth. Exp015 proved that our threshold removed useful training signal along with suspected noise.</p>
</section>
<section class="section interview-section" id="interview" data-track="Interview practice">
<div class="section-heading">
<p class="eyebrow">11 · Interview room</p>
<h2>Explain the project like you own it</h2>
<p>Say your answer aloud, then reveal the model answer. The phrasing is based on what we actually did.</p>
</div>
<div class="flashcard" id="flashcard">
<div class="flashcard-meta"><span id="flashcardNumber"></span><span>Tap reveal only after answering</span></div>
<h3 id="flashcardQuestion"></h3>
<div class="flashcard-answer" id="flashcardAnswer" hidden></div>
<div class="flashcard-actions">
<button class="primary-action small" id="revealFlashcard" type="button">Reveal answer</button>
<button class="text-action" id="nextFlashcard" type="button">Next question →</button>
</div>
</div>
</section>
<section class="section quiz-section" id="quiz" data-track="Knowledge check">
<div class="quiz-intro">
<p class="eyebrow">12 · Knowledge check</p>
<h2>Can you now reason about the lab?</h2>
<p>This is not trivia. Each question tests a decision you will face in another training project.</p>
<div class="quiz-progress"><i id="quizProgress"></i></div>
</div>
<div class="quiz-card" id="quizCard"></div>
</section>
<section class="section glossary-section" id="glossary" data-track="Glossary">
<div class="section-heading compact">
<p class="eyebrow">13 · Pocket glossary</p>
<h2>Terms you should be comfortable saying</h2>
</div>
<label class="search-box"><span>Search terms</span><input id="glossarySearch" type="search" placeholder="Try: epoch, logits, leakage…"></label>
<div class="glossary-grid" id="glossaryGrid"></div>
</section>
<section class="section mission-section" id="mission" data-track="Current mission">
<div class="mission-card">
<p class="eyebrow">Current lab checkpoint</p>
<div class="mission-grid">
<div><span class="mission-number">103</span><h2>Exp015 is closed</h2></div>
<div><p>The noise-pruned child reached 0.907792 validation accuracy versus Exp011's 0.929870. We reject the child, preserve Exp011, and do not open the locked test for a model that already lost on validation.</p><code>champion = exp-011 · next hypothesis not yet selected</code></div>
</div>
<p class="mission-rule"><strong>Scientific decision:</strong> a stable training run is not automatically a successful model. Exp015 completed correctly and still failed its promotion rule.</p>
</div>
</section>
</main>
<footer>
<div><strong>Open Model Training Lab</strong><p>Built from the project’s recorded experiments, not a generic ML tutorial.</p></div>
<div class="footer-progress"><span>Your reading progress</span><strong id="readingProgress">0%</strong><div><i id="readingProgressBar"></i></div></div>
</footer>
<script src="app.js"></script>
</body>
</html>
|