Download model_comparison_dashboard.html from mlengineer-ai/Jolt-4B: direct link, hf CLI and curl.
- Browser
- Download file 96.2 kB
-
https://huggingface.co/mlengineer-ai/Jolt-4B/resolve/main/model_comparison_dashboard.html
- Command line
-
hf download hf://mlengineer-ai/Jolt-4B/model_comparison_dashboard.html
-
curl -L -o model_comparison_dashboard.html https://huggingface.co/mlengineer-ai/Jolt-4B/resolve/main/model_comparison_dashboard.html
96.2 kB
| <html lang="en"> | |
| <head> | |
| <meta charset="utf-8"> | |
| <meta name="viewport" content="width=device-width,initial-scale=1"> | |
| <meta name="color-scheme" content="light"> | |
| <title>Model evaluation comparison</title> | |
| <style> | |
| :root{--ink:#172a36;--muted:#5d6d77;--line:#d9e2e5;--paper:#f4f7f7;--card:#fff;--teal:#087f78;--teal2:#26a69a;--blue:#3565a8;--orange:#c26b37;--purple:#7653a6;--gold:#b8871b;--red:#ac5144;--shadow:0 8px 24px rgba(25,54,65,.06)} | |
| *{box-sizing:border-box}body{margin:0;background:var(--paper);color:var(--ink);font:15px/1.5 system-ui,-apple-system,"Segoe UI",sans-serif}main{max-width:1440px;margin:auto;padding:32px 28px 64px}.eyebrow{color:var(--teal);font-size:12px;font-weight:750;letter-spacing:.12em;text-transform:uppercase}.hero{display:flex;justify-content:space-between;gap:32px;align-items:flex-end;margin-bottom:22px}.hero h1{font-size:clamp(30px,4vw,48px);line-height:1.07;margin:8px 0 12px;letter-spacing:-.035em}.hero p{max-width:790px;color:var(--muted);margin:0;font-size:16px}.stamp{font-size:12px;color:var(--muted);white-space:nowrap}.intro{padding:16px 18px;background:#e7f3f1;border-left:4px solid var(--teal);border-radius:10px;margin:18px 0 22px}.intro strong{color:#075d59}.controls{display:flex;align-items:end;flex-wrap:wrap;gap:14px;padding:16px;background:var(--card);border:1px solid var(--line);border-radius:14px;box-shadow:var(--shadow);margin-bottom:18px}.control{display:flex;flex-direction:column;gap:5px}.control label{font-size:11px;text-transform:uppercase;letter-spacing:.08em;font-weight:700;color:var(--muted)}select{font:inherit;color:var(--ink);background:#fff;border:1px solid #bac9ce;border-radius:8px;padding:9px 34px 9px 11px;min-width:170px}button{font:inherit;border:0;cursor:pointer}.scope-tabs{display:flex;gap:6px;flex-wrap:wrap}.scope-tabs button,.metric-toggle button{border:1px solid var(--line);background:#fff;border-radius:8px;padding:9px 12px;color:var(--muted)}.scope-tabs button[aria-pressed="true"],.metric-toggle button[aria-pressed="true"]{background:#163b48;color:#fff;border-color:#163b48}.metric-toggle{display:flex;gap:6px}.view-note{margin-left:auto;color:var(--muted);font-size:13px;max-width:370px}.kpis{display:grid;grid-template-columns:repeat(4,1fr);gap:14px;margin:16px 0 18px}.kpi,.panel{background:var(--card);border:1px solid var(--line);border-radius:14px;box-shadow:var(--shadow)}.kpi{padding:17px 18px;min-height:108px}.kpi-label{font-size:11px;text-transform:uppercase;letter-spacing:.09em;color:var(--muted);font-weight:700}.kpi-value{font-size:25px;font-weight:760;letter-spacing:-.025em;margin:5px 0 1px}.kpi-sub{font-size:12px;color:var(--muted)}.grid{display:grid;grid-template-columns:minmax(0,1.2fr) minmax(340px,.8fr);gap:16px;align-items:start}.panel{padding:20px 22px;margin-bottom:16px;min-width:0}.panel h2{font-size:19px;margin:0 0 5px;letter-spacing:-.015em}.panel>p,.panel .sub{color:var(--muted);font-size:13px;margin:0 0 15px}.panel-head{display:flex;justify-content:space-between;align-items:flex-start;gap:12px}.chart-scroll{overflow-x:auto}.chart{width:100%;min-width:440px;display:block}#datasetChart{width:100%;min-width:0}.legend{display:flex;flex-wrap:wrap;gap:12px;margin-top:8px;color:var(--muted);font-size:12px}.model-picker{position:relative;min-width:190px}.picker-details{position:relative;margin:0;border:0;background:transparent}.picker-details summary{display:block;cursor:pointer;padding:8px 28px 8px 10px;border:1px solid var(--line);border-radius:8px;background:#fff;font-size:12px;font-weight:650;white-space:nowrap}.picker-details[open] summary{border-color:var(--teal)}.model-picker-list{position:absolute;right:0;top:4px;z-index:10;display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:2px 10px;width:min(380px,80vw);top:calc(100% + 4px);max-height:260px;overflow:auto;border:1px solid var(--line);border-radius:8px;padding:7px 9px;background:#fff;box-shadow:var(--shadow)}.model-option{display:flex;align-items:center;gap:7px;min-width:0;padding:3px 0;color:var(--ink);font-size:11px}.model-option input{margin:0;accent-color:var(--teal)}.series-swatch{width:9px;height:9px;border-radius:50%;flex:none}.series-index{font-size:10px;color:var(--muted)}.legend span{display:flex;align-items:center;gap:5px}.dot{width:9px;height:9px;border-radius:50%;display:inline-block}.range-key{width:20px;height:2px;background:#233945;display:inline-block;position:relative}.range-key:before,.range-key:after{content:"";position:absolute;width:1px;height:7px;background:#233945;top:-3px}.range-key:before{left:0}.range-key:after{right:0}.badge{display:inline-block;padding:3px 7px;border-radius:99px;font-size:10px;line-height:1.3;font-weight:750;letter-spacing:.025em;vertical-align:middle}.badge.seed{background:#e8eff9;color:#365c8f}.badge.single{background:#edf0f1;color:#52636c}.badge.diag{background:#fff1d8;color:#81530a}.badge.exposure{background:#fae6e2;color:#913c33}.badge.clean{background:#e0f2ea;color:#166b46}.rank-item{cursor:pointer}.rank-item:hover text.rank-label{fill:var(--teal)}.table-wrap{overflow:auto;border:1px solid var(--line);border-radius:9px}.data-table{border-collapse:collapse;width:100%;font-size:12px;white-space:nowrap}.data-table th,.data-table td{padding:9px 10px;border-bottom:1px solid #e8edef;text-align:right}.data-table th{position:sticky;top:0;background:#f5f8f8;color:#52636c;font-size:10px;letter-spacing:.05em;text-transform:uppercase;cursor:pointer}.data-table th:first-child,.data-table td:first-child{text-align:left;position:sticky;left:0;background:#fff;z-index:1}.data-table th:first-child{background:#f5f8f8;z-index:2}.data-table tr:last-child td{border-bottom:0}.cell{min-width:66px;border-radius:5px;padding:3px 6px}.dataset-sub{font-size:10px;color:#71818a}.tiny{font-size:11px;color:var(--muted)}.insight{padding:14px 16px;background:#f7faf9;border:1px solid var(--line);border-radius:10px;font-size:14px;margin:12px 0}.insight strong{color:#135d58}.notes{display:grid;grid-template-columns:repeat(2,1fr);gap:12px}.note{padding:14px 15px;border:1px solid var(--line);border-radius:10px;background:#fbfcfc;font-size:13px}.note h3{font-size:14px;margin:0 0 5px}.note p{margin:0;color:var(--muted)}.exposure-list{display:grid;gap:8px}.exposure-row{padding:10px 12px;border-radius:8px;background:#f7f9f9;font-size:13px}.exposure-row p{margin:5px 0 0;color:var(--muted)}details{border:1px solid var(--line);border-radius:10px;margin:10px 0;background:#fff}summary{cursor:pointer;font-weight:700;padding:13px 15px}details .detail-body{padding:0 15px 15px}.diag-chart{max-height:440px;overflow:auto}.diag-bar-row{display:grid;grid-template-columns:minmax(120px,190px) 1fr 54px;gap:9px;align-items:center;margin:7px 0;font-size:12px}.diag-track{height:13px;background:#eef2f2;border-radius:99px;overflow:hidden}.diag-fill{height:100%;background:linear-gradient(90deg,#75b9b1,#087f78);border-radius:99px}.footer{margin-top:24px;padding-top:16px;border-top:1px solid var(--line);font-size:11px;color:var(--muted)}.footer a{color:var(--teal)}.sr-only{position:absolute;width:1px;height:1px;padding:0;margin:-1px;overflow:hidden;clip:rect(0,0,0,0);white-space:nowrap;border:0} | |
| @media(max-width:980px){main{padding:22px 16px 44px}.grid{grid-template-columns:1fr}.hero{display:block}.stamp{margin-top:10px}.kpis{grid-template-columns:repeat(2,1fr)}.view-note{margin-left:0;max-width:none;width:100%}} | |
| @media(max-width:580px){.model-picker{width:auto;min-width:170px}.model-picker-list{grid-template-columns:1fr}.kpis{grid-template-columns:1fr 1fr}.kpi{padding:13px;min-height:96px}.kpi-value{font-size:20px}.panel{padding:16px}.controls{align-items:stretch}.control,.control select{width:100%}.scope-tabs button{flex:1}.notes{grid-template-columns:1fr}} | |
| @media print{body{background:#fff}main{max-width:none;padding:0}.controls{display:none}.panel,.kpi{box-shadow:none;break-inside:avoid}.grid{grid-template-columns:1fr 1fr}details{break-inside:avoid}.footer{margin-top:12px}} | |
| </style> | |
| </head> | |
| <body> | |
| <main> | |
| <header class="hero"> | |
| <div><div class="eyebrow">Matched model evaluation</div><h1>Same test. Different models.</h1><p>Compare Jolt, Dohnuts, Laya, and Kev on the same frozen examples. Start with the shared text test, then switch to image results or the full 26-dataset view where coverage allows.</p></div> | |
| <div class="stamp">Built October 17, 2026 · Works offline</div> | |
| </header> | |
| <div class="intro"><strong>How to read this report.</strong> “Dohnuts 0.8B (retrained)” is the report name for Jolt’s B0 control. It was trained from the pinned Qwen base on Jolt’s training split with the Dohnuts recipe. “Dohnuts 0.8B (released)” is the separately trained downloaded checkpoint. The primary score is dataset-macro accuracy. Every dataset gets equal weight, so large image datasets do not drown out smaller tasks. Use row-weighted accuracy to see the result when every example counts equally. Results from different test suites appear below as diagnostics, not as part of the shared-test ranking.</div> | |
| <section class="panel" style="border-top:4px solid var(--orange)"><div class="panel-head"><div><div class="eyebrow">Separate benchmark comparison</div><h2>JevBench · public tasks</h2><p>Headline score is the mean of easy, standard, and hard accuracy, with each tier weighted equally. Pooled accuracy is shown beside it.</p></div><div class="tiny">231 public tasks · 303 official tasks unavailable</div></div><div id="jevbenchChart" class="diag-chart"></div><div class="table-wrap"><table id="jevbenchTable" class="data-table"></table></div><p class="tiny">This is a separate suite, so these scores do not enter the 26-dataset ranking above. Jolt and Dohnuts used a 4,096-token input limit; Laya used native 1,024-token truncation; Kev used its serving context. The table names each evaluation path.</p></section> | |
| <section class="controls" aria-label="Comparison filters"> | |
| <div class="control"><label>Evaluation lane</label><div class="scope-tabs" role="group" aria-label="Evaluation lane"><button data-scope="text" aria-pressed="true">Text · 18 datasets</button><button data-scope="image" aria-pressed="false">Images · 8 datasets</button><button data-scope="all" aria-pressed="false">All · 26 datasets</button></div></div> | |
| <div class="control"><label for="metric">Aggregate score</label><div class="metric-toggle" role="group"><button data-metric="macro" aria-pressed="true">Dataset macro</button><button data-metric="weighted" aria-pressed="false">Row weighted</button></div></div> | |
| <div class="view-note" id="laneNote"></div> | |
| </section> | |
| <section class="kpis" aria-label="Selected comparison summary"> | |
| <article class="kpi"><div class="kpi-label">Top score in view</div><div class="kpi-value" id="bestValue">—</div><div class="kpi-sub" id="bestName">—</div></article> | |
| <article class="kpi"><div class="kpi-label">Lead over next model</div><div class="kpi-value" id="leadValue">—</div><div class="kpi-sub" id="leadName">—</div></article> | |
| <article class="kpi"><div class="kpi-label">Shared evaluation data</div><div class="kpi-value" id="coverageValue">—</div><div class="kpi-sub" id="coverageName">—</div></article> | |
| <article class="kpi"><div class="kpi-label">Evidence per score</div><div class="kpi-value" id="runsValue">—</div><div class="kpi-sub" id="runsName">—</div></article> | |
| </section> | |
| <div class="grid"> | |
| <section class="panel"><div class="panel-head"><div><h2 id="rankTitle">Aggregate ranking</h2><p id="rankSub">Seed means are shown for Jolt runs; single-run results are labeled.</p></div></div><div class="chart-scroll"><svg id="ranking" class="chart" role="img" aria-label="Model accuracy comparison"></svg></div><div class="legend" id="familyLegend"></div><div class="tiny" style="margin-top:10px">Whiskers show the minimum and maximum among Jolt seeds. They describe run-to-run variation, not a confidence interval. Scores for single-run models have no whisker.</div><div class="insight" id="readingNote"></div></section> | |
| <section class="panel"><div class="panel-head"><div><h2>Dataset detail</h2><p>Select one or more models to compare their accuracy on each dataset in this lane.</p></div><div class="control model-picker"><label>Models to compare</label><details class="picker-details"><summary id="detailSummary">Choose models</summary><div id="detailModels" class="model-picker-list" role="group" aria-label="Models to compare"></div></details></div></div><div class="legend" id="detailLegend" aria-label="Selected model legend"></div><div class="chart-scroll"><svg id="datasetChart" class="chart" role="img" aria-label="Dataset accuracy comparison for selected models"></svg></div><div class="tiny" id="datasetCaption"></div></section> | |
| </div> | |
| <section class="panel"><div class="panel-head"><div><h2>Scores by dataset</h2><p>Cells show hard top-1 accuracy. Select a column heading to sort. The table follows the lane and model-family filters above.</p></div></div><div class="table-wrap"><table id="scoreTable" class="data-table"></table></div></section> | |
| <section class="panel"><h2>What the comparison says</h2><div id="interpretation" class="notes"></div></section> | |
| <section class="panel"><h2>Training-data exposure and limits</h2><p>Scoring every model on the same test rows makes the inputs comparable. It does not make their training histories equally clean.</p><div class="exposure-list" id="exposure"></div></section> | |
| <section class="panel"><h2>Other evaluations</h2><p>These results use other splits or task suites. They help describe each model, but they do not belong in the matched test ranking.</p> | |
| <details open><summary>Jolt training experiments · development screens</summary><div class="detail-body"><p class="sub">The first chart shows capped development results at updates 400 and 800. Full-development scores appear only when that evaluation was run. These are training-selection measurements, not locked-test scores.</p><div id="screenChart" class="diag-chart"></div><div class="table-wrap"><table id="screenTable" class="data-table"></table></div></div></details> | |
| <details><summary>Kev official development suites</summary><div class="detail-body"><p class="sub">Kev's own development partitions, evaluated with its native benchmark code. These suites differ from the Jolt frozen test.</p><div class="table-wrap"><table id="kevTable" class="data-table"></table></div></div></details> | |
| <details><summary>Jolt 0.8B L2 secondary suites</summary><div class="detail-body"><p class="sub">The additional application collection has its own task mix. Read their scores within each suite only.</p><div class="table-wrap"><table id="secondaryTable" class="data-table"></table></div></div></details> | |
| </section> | |
| <footer class="footer"><div>Primary evaluation source: Jolt frozen test split, SHA-256 <code>fdac64d536d67c011d55d00352f0be71fa3bdfb7811c0be0e05a5d66cec98095</code>. All included predictions were checked for exact test IDs, dataset names, and target labels. Accuracy uses top-1 label agreement. Positive temperature scaling does not change top-1 predictions, so every family uses the same accuracy rule.</div><div style="margin-top:8px">Source notes: <a href="dohnuts/experiment_findings_report.md">Jolt experiment report</a> · Kev local comparison report is stored at <code>/mnt/990Pro/kev-evaluation/comparison_report.md</code>. The dashboard embeds preaggregated scores and does not load files or libraries from the network.</div></footer> | |
| </main> | |
| <script> | |
| const DATA={"meta":{"test_file":"dohnuts/data/processed/v1/test.jsonl","test_sha256":"fdac64d536d67c011d55d00352f0be71fa3bdfb7811c0be0e05a5d66cec98095","rows":180031,"dataset_count":26,"scope_counts":{"text":{"datasets":18,"rows":56426},"image":{"datasets":8,"rows":123605},"all":{"datasets":26,"rows":180031}},"metric":"Hard top-1 accuracy: predicted candidate index matches target argmax. Dataset macro averages dataset accuracies equally; row-weighted pools rows.","models":11},"datasets":{"ag_news":{"rows":7600,"kind":"text"},"aokvqa":{"rows":1138,"kind":"image"},"banking77":{"rows":3080,"kind":"text"},"boolq":{"rows":3270,"kind":"text"},"clevr_attribute":{"rows":53734,"kind":"image"},"clevr_count":{"rows":35422,"kind":"image"},"clevr_exist":{"rows":20196,"kind":"image"},"contract_nli":{"rows":1173,"kind":"text"},"emotion":{"rows":2000,"kind":"text"},"esci_es":{"rows":1482,"kind":"text"},"esci_jp":{"rows":1633,"kind":"text"},"esci_us":{"rows":1145,"kind":"text"},"mail_phishing":{"rows":1050,"kind":"text"},"mail_spam":{"rows":854,"kind":"text"},"massive_en-US":{"rows":2970,"kind":"text"},"massive_zh-CN":{"rows":2921,"kind":"text"},"scienceqa":{"rows":2017,"kind":"image"},"screenqa_choice":{"rows":848,"kind":"image"},"screenqa_noul":{"rows":2148,"kind":"image"},"sharc":{"rows":8276,"kind":"text"},"sms_spam":{"rows":794,"kind":"text"},"typed_decisions":{"rows":2000,"kind":"text"},"vqav2_yesno":{"rows":8102,"kind":"image"},"wikiqa":{"rows":6160,"kind":"text"},"xnli_en":{"rows":5009,"kind":"text"},"xnli_zh":{"rows":5009,"kind":"text"}},"models":[{"id":"jolt_b0","name":"Dohnuts 0.8B (retrained)","family":"Jolt","base":"Qwen3.5-0.8B","note":"Jolt baseline trained from the pinned Qwen base with the Dohnuts RLCD recipe, using Jolt train rows. Test rows were excluded from campaign training, selection, and calibration.","run_count":3,"runs":[{"seed":"42","scope":{}},{"seed":"43","scope":{}},{"seed":"44","scope":{}}],"scores":{"text":{"macro":0.7838329041511867,"weighted":0.7901440234407305,"macro_sd":0.011511660568638176,"macro_min":0.7706281807674302,"macro_max":0.7917561705770998,"weighted_sd":0.01137007534084037,"rows":56426,"datasets":18,"run_count":3,"weighted_min":0.7770176868819338,"weighted_max":0.7969375819657605},"image":{"macro":0.7702316778458088,"weighted":0.9151113088736972,"macro_sd":0.0017359305961422468,"macro_min":0.7682284199723828,"macro_max":0.7712939158085842,"weighted_sd":0.0016333233449448451,"rows":123605,"datasets":8,"run_count":3,"weighted_min":0.9139921524210186,"weighted_max":0.9169855588366166},"all":{"macro":0.7796479114418396,"weighted":0.8759435874932651,"macro_sd":0.007755423280125163,"macro_min":0.7708330223185544,"macro_max":0.7854227943245952,"weighted_sd":0.002456433063757501,"rows":180031,"datasets":26,"run_count":3,"weighted_min":0.8731162966378012,"weighted_max":0.8775544211830185}},"by_dataset":{"sharc":{"rows":8276,"kind":"text","mean":0.6808442081520863,"runs":[0.6764137264378927,0.6785886901884969,0.6875302078298695],"min":0.6764137264378927,"max":0.6875302078298695,"run_count":3},"contract_nli":{"rows":1173,"kind":"text","mean":0.8249502699630576,"runs":[0.8218243819266837,0.8235294117647058,0.8294970161977835],"min":0.8218243819266837,"max":0.8294970161977835,"run_count":3},"wikiqa":{"rows":6160,"kind":"text","mean":0.9582792207792208,"runs":[0.9607142857142857,0.9586038961038961,0.9555194805194805],"min":0.9555194805194805,"max":0.9607142857142857,"run_count":3},"esci_us":{"rows":1145,"kind":"text","mean":0.520524017467249,"runs":[0.5432314410480349,0.5135371179039301,0.5048034934497817],"min":0.5048034934497817,"max":0.5432314410480349,"run_count":3},"esci_es":{"rows":1482,"kind":"text","mean":0.5636527215474584,"runs":[0.5695006747638327,0.5769230769230769,0.5445344129554656],"min":0.5445344129554656,"max":0.5769230769230769,"run_count":3},"esci_jp":{"rows":1633,"kind":"text","mean":0.6125739946927945,"runs":[0.6080832823025107,0.6117575015309247,0.617881200244948],"min":0.6080832823025107,"max":0.617881200244948,"run_count":3},"mail_spam":{"rows":854,"kind":"text","mean":0.9855581576893052,"runs":[0.9882903981264637,0.9859484777517564,0.9824355971896955],"min":0.9824355971896955,"max":0.9882903981264637,"run_count":3},"mail_phishing":{"rows":1050,"kind":"text","mean":0.9907936507936507,"runs":[0.9933333333333333,0.9914285714285714,0.9876190476190476],"min":0.9876190476190476,"max":0.9933333333333333,"run_count":3},"sms_spam":{"rows":794,"kind":"text","mean":0.9937027707808563,"runs":[0.9949622166246851,0.9924433249370277,0.9937027707808564],"min":0.9924433249370277,"max":0.9949622166246851,"run_count":3},"massive_en-US":{"rows":2970,"kind":"text","mean":0.7602693602693602,"runs":[0.7885521885521886,0.6831649831649832,0.8090909090909091],"min":0.6831649831649832,"max":0.8090909090909091,"run_count":3},"massive_zh-CN":{"rows":2921,"kind":"text","mean":0.744151546274107,"runs":[0.764464224580623,0.6737418692228688,0.7942485450188291],"min":0.6737418692228688,"max":0.7942485450188291,"run_count":3},"boolq":{"rows":3270,"kind":"text","mean":0.8265035677879715,"runs":[0.8324159021406727,0.8287461773700305,0.818348623853211],"min":0.818348623853211,"max":0.8324159021406727,"run_count":3},"ag_news":{"rows":7600,"kind":"text","mean":0.8959649122807017,"runs":[0.896578947368421,0.8892105263157895,0.9021052631578947],"min":0.8892105263157895,"max":0.9021052631578947,"run_count":3},"emotion":{"rows":2000,"kind":"text","mean":0.758,"runs":[0.774,0.7395,0.7605],"min":0.7395,"max":0.774,"run_count":3},"banking77":{"rows":3080,"kind":"text","mean":0.7093073593073593,"runs":[0.737012987012987,0.6503246753246753,0.7405844155844156],"min":0.6503246753246753,"max":0.7405844155844156,"run_count":3},"xnli_en":{"rows":5009,"kind":"text","mean":0.835296466360551,"runs":[0.8400878418846077,0.8334997005390298,0.8323018566580156],"min":0.8323018566580156,"max":0.8400878418846077,"run_count":3},"xnli_zh":{"rows":5009,"kind":"text","mean":0.7481200505756305,"runs":[0.752645238570573,0.7448592533439808,0.7468556598123378],"min":0.7448592533439808,"max":0.752645238570573,"run_count":3},"typed_decisions":{"rows":2000,"kind":"text","mean":0.7005,"runs":[0.7095,0.6955,0.6965],"min":0.6955,"max":0.7095,"run_count":3},"aokvqa":{"rows":1138,"kind":"image","mean":0.7779730521382543,"runs":[0.7741652021089631,0.7776801405975395,0.7820738137082601],"min":0.7741652021089631,"max":0.7820738137082601,"run_count":3},"scienceqa":{"rows":2017,"kind":"image","mean":0.8517600396628656,"runs":[0.8517600396628656,0.8517600396628656,0.8517600396628656],"min":0.8517600396628656,"max":0.8517600396628656,"run_count":3},"vqav2_yesno":{"rows":8102,"kind":"image","mean":0.812474286184481,"runs":[0.8120217230313502,0.8147371019501358,0.8106640335719575],"min":0.8106640335719575,"max":0.8147371019501358,"run_count":3},"clevr_exist":{"rows":20196,"kind":"image","mean":0.9760678682247309,"runs":[0.9771736977619331,0.9755397108338285,0.9754901960784313],"min":0.9754901960784313,"max":0.9771736977619331,"run_count":3},"clevr_attribute":{"rows":53734,"kind":"image","mean":0.9839021848364164,"runs":[0.9852235083931962,0.9833066587263185,0.9831763873897347],"min":0.9831763873897347,"max":0.9852235083931962,"run_count":3},"clevr_count":{"rows":35422,"kind":"image","mean":0.836683417085427,"runs":[0.8297103495003105,0.8441081813562193,0.8362317203997516],"min":0.8297103495003105,"max":0.8441081813562193,"run_count":3},"screenqa_choice":{"rows":848,"kind":"image","mean":0.21147798742138366,"runs":[0.2205188679245283,0.2169811320754717,0.19693396226415094],"min":0.19693396226415094,"max":0.2205188679245283,"run_count":3},"screenqa_noul":{"rows":2148,"kind":"image","mean":0.7115145872129113,"runs":[0.7188081936685289,0.7062383612662942,0.7094972067039106],"min":0.7062383612662942,"max":0.7188081936685289,"run_count":3}}},{"id":"jolt_l2_08","name":"Jolt 0.8B L2","family":"Jolt","base":"Qwen3.5-0.8B","note":"Jolt direct-composite objective, three matched seeds.","run_count":3,"runs":[{"seed":"42","scope":{}},{"seed":"43","scope":{}},{"seed":"44","scope":{}}],"scores":{"text":{"macro":0.8221927952803824,"weighted":0.8253701012535593,"macro_sd":0.0016162635989367905,"macro_min":0.8207989975658042,"macro_max":0.8239645448107306,"weighted_sd":0.0013302767595242486,"rows":56426,"datasets":18,"run_count":3,"weighted_min":0.824584411441534,"weighted_max":0.8269060362244356},"image":{"macro":0.7915966459326963,"weighted":0.9319175330016316,"macro_sd":0.0027212752674294688,"macro_min":0.7897190431138774,"macro_max":0.7947174864852459,"weighted_sd":0.0003808741518926863,"rows":123605,"datasets":8,"run_count":3,"weighted_min":0.931483354233243,"weighted_max":0.9321952995428987},"all":{"macro":0.8127785954810944,"weighted":0.8985230321444639,"macro_sd":0.0006465454546948455,"macro_min":0.8121344018444133,"macro_max":0.813427467365545,"weighted_sd":0.00016279808354953474,"rows":180031,"datasets":26,"run_count":3,"weighted_min":0.8983952763690698,"weighted_max":0.8987063339091601}},"by_dataset":{"sharc":{"rows":8276,"kind":"text","mean":0.718422748509747,"runs":[0.7266795553407444,0.7135089415176413,0.7150797486708554],"min":0.7135089415176413,"max":0.7266795553407444,"run_count":3},"contract_nli":{"rows":1173,"kind":"text","mean":0.8570616652458085,"runs":[0.8593350383631714,0.8499573742540495,0.8618925831202046],"min":0.8499573742540495,"max":0.8618925831202046,"run_count":3},"wikiqa":{"rows":6160,"kind":"text","mean":0.9580086580086581,"runs":[0.9577922077922078,0.9581168831168831,0.9581168831168831],"min":0.9577922077922078,"max":0.9581168831168831,"run_count":3},"esci_us":{"rows":1145,"kind":"text","mean":0.5737991266375545,"runs":[0.588646288209607,0.5668122270742358,0.5659388646288209],"min":0.5659388646288209,"max":0.588646288209607,"run_count":3},"esci_es":{"rows":1482,"kind":"text","mean":0.6084120557804769,"runs":[0.6126855600539811,0.6126855600539811,0.5998650472334683],"min":0.5998650472334683,"max":0.6126855600539811,"run_count":3},"esci_jp":{"rows":1633,"kind":"text","mean":0.6319657072872015,"runs":[0.6301285976729945,0.646050214329455,0.6197183098591549],"min":0.6197183098591549,"max":0.646050214329455,"run_count":3},"mail_spam":{"rows":854,"kind":"text","mean":0.9867291178766587,"runs":[0.9859484777517564,0.9882903981264637,0.9859484777517564],"min":0.9859484777517564,"max":0.9882903981264637,"run_count":3},"mail_phishing":{"rows":1050,"kind":"text","mean":0.9926984126984127,"runs":[0.9923809523809524,0.9933333333333333,0.9923809523809524],"min":0.9923809523809524,"max":0.9933333333333333,"run_count":3},"sms_spam":{"rows":794,"kind":"text","mean":0.9941225860621327,"runs":[0.9937027707808564,0.9949622166246851,0.9937027707808564],"min":0.9937027707808564,"max":0.9949622166246851,"run_count":3},"massive_en-US":{"rows":2970,"kind":"text","mean":0.8613916947250281,"runs":[0.8632996632996633,0.8548821548821549,0.865993265993266],"min":0.8548821548821549,"max":0.865993265993266,"run_count":3},"massive_zh-CN":{"rows":2921,"kind":"text","mean":0.8419491041880635,"runs":[0.8418349880178021,0.8404655939746662,0.843546730571722],"min":0.8404655939746662,"max":0.843546730571722,"run_count":3},"boolq":{"rows":3270,"kind":"text","mean":0.8372069317023446,"runs":[0.8360856269113149,0.8385321100917431,0.8370030581039756],"min":0.8360856269113149,"max":0.8385321100917431,"run_count":3},"ag_news":{"rows":7600,"kind":"text","mean":0.9010087719298245,"runs":[0.8989473684210526,0.9025,0.901578947368421],"min":0.8989473684210526,"max":0.9025,"run_count":3},"emotion":{"rows":2000,"kind":"text","mean":0.8401666666666667,"runs":[0.8465,0.8375,0.8365],"min":0.8365,"max":0.8465,"run_count":3},"banking77":{"rows":3080,"kind":"text","mean":0.8461038961038961,"runs":[0.8474025974025974,0.8483766233766233,0.8425324675324676],"min":0.8425324675324676,"max":0.8483766233766233,"run_count":3},"xnli_en":{"rows":5009,"kind":"text","mean":0.8512011712251281,"runs":[0.8500698742263925,0.8500698742263925,0.8534637652225994],"min":0.8500698742263925,"max":0.8534637652225994,"run_count":3},"xnli_zh":{"rows":5009,"kind":"text","mean":0.7655553337326145,"runs":[0.7654222399680575,0.7646236773807147,0.7666200838490717],"min":0.7646236773807147,"max":0.7666200838490717,"run_count":3},"typed_decisions":{"rows":2000,"kind":"text","mean":0.7336666666666667,"runs":[0.7345,0.732,0.7345],"min":0.732,"max":0.7345,"run_count":3},"aokvqa":{"rows":1138,"kind":"image","mean":0.7867603983596952,"runs":[0.7855887521968365,0.7838312829525483,0.7908611599297012],"min":0.7838312829525483,"max":0.7908611599297012,"run_count":3},"scienceqa":{"rows":2017,"kind":"image","mean":0.8958849776896382,"runs":[0.8889439762022806,0.90233019335647,0.8963807635101636],"min":0.8889439762022806,"max":0.90233019335647,"run_count":3},"vqav2_yesno":{"rows":8102,"kind":"image","mean":0.8122685756603308,"runs":[0.8143668230066651,0.8131325598617626,0.8093063441125647],"min":0.8093063441125647,"max":0.8143668230066651,"run_count":3},"clevr_exist":{"rows":20196,"kind":"image","mean":0.9791542879778175,"runs":[0.9787086551792434,0.9788076846900376,0.9799465240641712],"min":0.9787086551792434,"max":0.9799465240641712,"run_count":3},"clevr_attribute":{"rows":53734,"kind":"image","mean":0.9879033758886365,"runs":[0.987847545315815,0.9885361223806156,0.9873264599694793],"min":0.9873264599694793,"max":0.9885361223806156,"run_count":3},"clevr_count":{"rows":35422,"kind":"image","mean":0.8824929892910244,"runs":[0.8815707752244368,0.8827564790243352,0.8831517136243012],"min":0.8815707752244368,"max":0.8831517136243012,"run_count":3},"screenqa_choice":{"rows":848,"kind":"image","mean":0.2578616352201258,"runs":[0.2535377358490566,0.24764150943396226,0.27240566037735847],"min":0.24764150943396226,"max":0.27240566037735847,"run_count":3},"screenqa_noul":{"rows":2148,"kind":"image","mean":0.7304469273743016,"runs":[0.7271880819366853,0.7257914338919925,0.7383612662942272],"min":0.7257914338919925,"max":0.7383612662942272,"run_count":3}}},{"id":"jolt_l2_2b","name":"Jolt 2B L2","family":"Jolt","base":"Qwen3.5-2B","note":"Jolt 2B transfer using the L2 objective, three matched seeds.","run_count":3,"runs":[{"seed":"42","scope":{}},{"seed":"43","scope":{}},{"seed":"44","scope":{}}],"scores":{"text":{"macro":0.8351972018453758,"weighted":0.8411902314535853,"macro_sd":0.001377254530902626,"macro_min":0.8338596432913975,"macro_max":0.8366109864101995,"weighted_sd":0.0010796082947944399,"rows":56426,"datasets":18,"run_count":3,"weighted_min":0.8399496685924929,"weighted_max":0.8419168468436536},"image":{"macro":0.8191955343382537,"weighted":0.952941493736769,"macro_sd":0.00554670635349887,"macro_min":0.8132689854852595,"macro_max":0.8242618934971736,"weighted_sd":0.0005798164883094751,"rows":123605,"datasets":8,"run_count":3,"weighted_min":0.9524857408680879,"weighted_max":0.9535941102706201},"all":{"macro":0.8302736118431843,"weighted":0.9179159885427141,"macro_sd":0.001653374289655098,"macro_min":0.8283972864962931,"macro_max":0.8315170595247005,"weighted_sd":0.0005713501968624108,"rows":180031,"datasets":26,"run_count":3,"weighted_min":0.9173920047102999,"weighted_max":0.9185251428920574}},"by_dataset":{"sharc":{"rows":8276,"kind":"text","mean":0.750362493958434,"runs":[0.7533832769453842,0.7415418076365394,0.7561623972933784],"min":0.7415418076365394,"max":0.7561623972933784,"run_count":3},"contract_nli":{"rows":1173,"kind":"text","mean":0.8607558965615231,"runs":[0.8576300085251491,0.8627450980392157,0.8618925831202046],"min":0.8576300085251491,"max":0.8627450980392157,"run_count":3},"wikiqa":{"rows":6160,"kind":"text","mean":0.9616341991341991,"runs":[0.962012987012987,0.9618506493506493,0.961038961038961],"min":0.961038961038961,"max":0.962012987012987,"run_count":3},"esci_us":{"rows":1145,"kind":"text","mean":0.5895196506550219,"runs":[0.5877729257641922,0.6008733624454149,0.5799126637554585],"min":0.5799126637554585,"max":0.6008733624454149,"run_count":3},"esci_es":{"rows":1482,"kind":"text","mean":0.6401259559154296,"runs":[0.6322537112010796,0.6545209176788124,0.6336032388663968],"min":0.6322537112010796,"max":0.6545209176788124,"run_count":3},"esci_jp":{"rows":1633,"kind":"text","mean":0.6315574607062665,"runs":[0.6325780771586038,0.6362522963870177,0.6258420085731782],"min":0.6258420085731782,"max":0.6362522963870177,"run_count":3},"mail_spam":{"rows":854,"kind":"text","mean":0.9898516783762684,"runs":[0.9894613583138173,0.990632318501171,0.9894613583138173],"min":0.9894613583138173,"max":0.990632318501171,"run_count":3},"mail_phishing":{"rows":1050,"kind":"text","mean":0.9920634920634921,"runs":[0.9914285714285714,0.9942857142857143,0.9904761904761905],"min":0.9904761904761905,"max":0.9942857142857143,"run_count":3},"sms_spam":{"rows":794,"kind":"text","mean":0.9953820319059613,"runs":[0.9937027707808564,0.9974811083123426,0.9949622166246851],"min":0.9937027707808564,"max":0.9974811083123426,"run_count":3},"massive_en-US":{"rows":2970,"kind":"text","mean":0.8723905723905724,"runs":[0.8703703703703703,0.8723905723905724,0.8744107744107744],"min":0.8703703703703703,"max":0.8744107744107744,"run_count":3},"massive_zh-CN":{"rows":2921,"kind":"text","mean":0.8493666552550496,"runs":[0.8497090037658336,0.8476549127011298,0.8507360492981856],"min":0.8476549127011298,"max":0.8507360492981856,"run_count":3},"boolq":{"rows":3270,"kind":"text","mean":0.8760448521916411,"runs":[0.8776758409785933,0.8761467889908257,0.8743119266055046],"min":0.8743119266055046,"max":0.8776758409785933,"run_count":3},"ag_news":{"rows":7600,"kind":"text","mean":0.9042543859649124,"runs":[0.9047368421052632,0.9042105263157895,0.9038157894736842],"min":0.9038157894736842,"max":0.9047368421052632,"run_count":3},"emotion":{"rows":2000,"kind":"text","mean":0.8330000000000001,"runs":[0.841,0.84,0.818],"min":0.818,"max":0.841,"run_count":3},"banking77":{"rows":3080,"kind":"text","mean":0.8658008658008658,"runs":[0.8600649350649351,0.8707792207792208,0.8665584415584415],"min":0.8600649350649351,"max":0.8707792207792208,"run_count":3},"xnli_en":{"rows":5009,"kind":"text","mean":0.8734943767884474,"runs":[0.8762227989618686,0.8684368137352765,0.8758235176681972],"min":0.8684368137352765,"max":0.8762227989618686,"run_count":3},"xnli_zh":{"rows":5009,"kind":"text","mean":0.7897783988820124,"runs":[0.7921740866440408,0.7801956478338989,0.7969654621680974],"min":0.7801956478338989,"max":0.7969654621680974,"run_count":3},"typed_decisions":{"rows":2000,"kind":"text","mean":0.7581666666666668,"runs":[0.76,0.759,0.7555],"min":0.7555,"max":0.76,"run_count":3},"aokvqa":{"rows":1138,"kind":"image","mean":0.8482718219097832,"runs":[0.8409490333919156,0.8523725834797891,0.851493848857645],"min":0.8409490333919156,"max":0.8523725834797891,"run_count":3},"scienceqa":{"rows":2017,"kind":"image","mean":0.9416625351181622,"runs":[0.9400099157164105,0.9424888448190382,0.9424888448190382],"min":0.9400099157164105,"max":0.9424888448190382,"run_count":3},"vqav2_yesno":{"rows":8102,"kind":"image","mean":0.858430017279684,"runs":[0.8597877067390768,0.8583065909651938,0.8571957541347816],"min":0.8571957541347816,"max":0.8597877067390768,"run_count":3},"clevr_exist":{"rows":20196,"kind":"image","mean":0.9887106357694592,"runs":[0.9894038423450188,0.9872251931075461,0.989502871855813],"min":0.9872251931075461,"max":0.989502871855813,"run_count":3},"clevr_attribute":{"rows":53734,"kind":"image","mean":0.9938090098137741,"runs":[0.9930397885882309,0.9941191796627833,0.9942680611903079],"min":0.9930397885882309,"max":0.9942680611903079,"run_count":3},"clevr_count":{"rows":35422,"kind":"image","mean":0.9259405642444433,"runs":[0.9263169781491728,0.9254700468635312,0.9260346677206256],"min":0.9254700468635312,"max":0.9263169781491728,"run_count":3},"screenqa_choice":{"rows":848,"kind":"image","mean":0.26179245283018865,"runs":[0.2275943396226415,0.267688679245283,0.29009433962264153],"min":0.2275943396226415,"max":0.29009433962264153,"run_count":3},"screenqa_noul":{"rows":2148,"kind":"image","mean":0.7349472377405338,"runs":[0.729050279329609,0.7327746741154563,0.7430167597765364],"min":0.729050279329609,"max":0.7430167597765364,"run_count":3}}},{"id":"jolt_l2_4b","name":"Jolt 4B L2","family":"Jolt","base":"Qwen3.5-4B","note":"Jolt L2; BF16 base for dashboard test evaluation; seed 42, step 3600. One checkpoint evaluated on the frozen test split.","run_count":1,"runs":[{"seed":"42","scope":{}}],"scores":{"text":{"macro":0.8511217479523615,"weighted":0.859107503633077,"macro_sd":null,"macro_min":0.8511217479523615,"macro_max":0.8511217479523615,"weighted_sd":null,"rows":56426,"datasets":18,"run_count":1,"weighted_min":0.859107503633077,"weighted_max":0.859107503633077},"image":{"macro":0.8396383485107799,"weighted":0.9692488167954371,"macro_sd":null,"macro_min":0.8396383485107799,"macro_max":0.8396383485107799,"weighted_sd":null,"rows":123605,"datasets":8,"run_count":1,"weighted_min":0.9692488167954371,"weighted_max":0.9692488167954371},"all":{"macro":0.8475883942780288,"weighted":0.9347279079714049,"macro_sd":null,"macro_min":0.8475883942780288,"macro_max":0.8475883942780288,"weighted_sd":null,"rows":180031,"datasets":26,"run_count":1,"weighted_min":0.9347279079714049,"weighted_max":0.9347279079714049}},"by_dataset":{"ag_news":{"rows":7600,"kind":"text","mean":0.9126315789473685,"runs":[0.9126315789473685],"min":0.9126315789473685,"max":0.9126315789473685,"run_count":1},"aokvqa":{"rows":1138,"kind":"image","mean":0.8725834797891037,"runs":[0.8725834797891037],"min":0.8725834797891037,"max":0.8725834797891037,"run_count":1},"banking77":{"rows":3080,"kind":"text","mean":0.8935064935064935,"runs":[0.8935064935064935],"min":0.8935064935064935,"max":0.8935064935064935,"run_count":1},"boolq":{"rows":3270,"kind":"text","mean":0.9073394495412844,"runs":[0.9073394495412844],"min":0.9073394495412844,"max":0.9073394495412844,"run_count":1},"clevr_attribute":{"rows":53734,"kind":"image","mean":0.9987903375888637,"runs":[0.9987903375888637],"min":0.9987903375888637,"max":0.9987903375888637,"run_count":1},"clevr_count":{"rows":35422,"kind":"image","mean":0.9630455649031675,"runs":[0.9630455649031675],"min":0.9630455649031675,"max":0.9630455649031675,"run_count":1},"clevr_exist":{"rows":20196,"kind":"image","mean":0.9947019211725094,"runs":[0.9947019211725094],"min":0.9947019211725094,"max":0.9947019211725094,"run_count":1},"contract_nli":{"rows":1173,"kind":"text","mean":0.8797953964194374,"runs":[0.8797953964194374],"min":0.8797953964194374,"max":0.8797953964194374,"run_count":1},"emotion":{"rows":2000,"kind":"text","mean":0.85,"runs":[0.85],"min":0.85,"max":0.85,"run_count":1},"esci_es":{"rows":1482,"kind":"text","mean":0.6538461538461539,"runs":[0.6538461538461539],"min":0.6538461538461539,"max":0.6538461538461539,"run_count":1},"esci_jp":{"rows":1633,"kind":"text","mean":0.6466625842008573,"runs":[0.6466625842008573],"min":0.6466625842008573,"max":0.6466625842008573,"run_count":1},"esci_us":{"rows":1145,"kind":"text","mean":0.6096069868995633,"runs":[0.6096069868995633],"min":0.6096069868995633,"max":0.6096069868995633,"run_count":1},"mail_phishing":{"rows":1050,"kind":"text","mean":0.9904761904761905,"runs":[0.9904761904761905],"min":0.9904761904761905,"max":0.9904761904761905,"run_count":1},"mail_spam":{"rows":854,"kind":"text","mean":0.990632318501171,"runs":[0.990632318501171],"min":0.990632318501171,"max":0.990632318501171,"run_count":1},"massive_en-US":{"rows":2970,"kind":"text","mean":0.8892255892255893,"runs":[0.8892255892255893],"min":0.8892255892255893,"max":0.8892255892255893,"run_count":1},"massive_zh-CN":{"rows":2921,"kind":"text","mean":0.860664156110921,"runs":[0.860664156110921],"min":0.860664156110921,"max":0.860664156110921,"run_count":1},"scienceqa":{"rows":2017,"kind":"image","mean":0.9752107089737233,"runs":[0.9752107089737233],"min":0.9752107089737233,"max":0.9752107089737233,"run_count":1},"screenqa_choice":{"rows":848,"kind":"image","mean":0.2841981132075472,"runs":[0.2841981132075472],"min":0.2841981132075472,"max":0.2841981132075472,"run_count":1},"screenqa_noul":{"rows":2148,"kind":"image","mean":0.749534450651769,"runs":[0.749534450651769],"min":0.749534450651769,"max":0.749534450651769,"run_count":1},"sharc":{"rows":8276,"kind":"text","mean":0.7664330594490092,"runs":[0.7664330594490092],"min":0.7664330594490092,"max":0.7664330594490092,"run_count":1},"sms_spam":{"rows":794,"kind":"text","mean":0.9937027707808564,"runs":[0.9937027707808564],"min":0.9937027707808564,"max":0.9937027707808564,"run_count":1},"typed_decisions":{"rows":2000,"kind":"text","mean":0.773,"runs":[0.773],"min":0.773,"max":0.773,"run_count":1},"vqav2_yesno":{"rows":8102,"kind":"image","mean":0.8790422117995557,"runs":[0.8790422117995557],"min":0.8790422117995557,"max":0.8790422117995557,"run_count":1},"wikiqa":{"rows":6160,"kind":"text","mean":0.9637987012987013,"runs":[0.9637987012987013],"min":0.9637987012987013,"max":0.9637987012987013,"run_count":1},"xnli_en":{"rows":5009,"kind":"text","mean":0.9057696146935517,"runs":[0.9057696146935517],"min":0.9057696146935517,"max":0.9057696146935517,"run_count":1},"xnli_zh":{"rows":5009,"kind":"text","mean":0.8331004192453584,"runs":[0.8331004192453584],"min":0.8331004192453584,"max":0.8331004192453584,"run_count":1}}},{"id":"jolt_l2_9b_bf16","name":"Jolt 9B L2","family":"Jolt","base":"Qwen3.5-9B (BF16)","note":"Single seed-42 run, selected at step 3,600 by development macro accuracy. Trained and evaluated with BF16 base weights and compute. The 180,031 locked test rows were held out from training, checkpoint selection, and calibration.","run_count":1,"runs":[{"seed":"42","scope":{}}],"scores":{"text":{"macro":0.8580750720284059,"weighted":0.8655407081841704,"macro_sd":null,"macro_min":0.8580750720284059,"macro_max":0.8580750720284059,"weighted_sd":null,"rows":56426,"datasets":18,"run_count":1,"weighted_min":0.8655407081841704,"weighted_max":0.8655407081841704},"image":{"macro":0.8487683328967103,"weighted":0.9751385461753165,"macro_sd":null,"macro_min":0.8487683328967103,"macro_max":0.8487683328967103,"weighted_sd":null,"rows":123605,"datasets":8,"run_count":1,"weighted_min":0.9751385461753165,"weighted_max":0.9751385461753165},"all":{"macro":0.8552114599878842,"weighted":0.940787975404236,"macro_sd":null,"macro_min":0.8552114599878842,"macro_max":0.8552114599878842,"weighted_sd":null,"rows":180031,"datasets":26,"run_count":1,"weighted_min":0.940787975404236,"weighted_max":0.940787975404236}},"by_dataset":{"ag_news":{"rows":7600,"kind":"text","mean":0.9126315789473685,"runs":[0.9126315789473685],"min":0.9126315789473685,"max":0.9126315789473685,"run_count":1},"banking77":{"rows":3080,"kind":"text","mean":0.8948051948051948,"runs":[0.8948051948051948],"min":0.8948051948051948,"max":0.8948051948051948,"run_count":1},"boolq":{"rows":3270,"kind":"text","mean":0.9110091743119266,"runs":[0.9110091743119266],"min":0.9110091743119266,"max":0.9110091743119266,"run_count":1},"contract_nli":{"rows":1173,"kind":"text","mean":0.875532821824382,"runs":[0.875532821824382],"min":0.875532821824382,"max":0.875532821824382,"run_count":1},"emotion":{"rows":2000,"kind":"text","mean":0.8595,"runs":[0.8595],"min":0.8595,"max":0.8595,"run_count":1},"esci_es":{"rows":1482,"kind":"text","mean":0.6612685560053981,"runs":[0.6612685560053981],"min":0.6612685560053981,"max":0.6612685560053981,"run_count":1},"esci_jp":{"rows":1633,"kind":"text","mean":0.653398652786283,"runs":[0.653398652786283],"min":0.653398652786283,"max":0.653398652786283,"run_count":1},"esci_us":{"rows":1145,"kind":"text","mean":0.6524017467248908,"runs":[0.6524017467248908],"min":0.6524017467248908,"max":0.6524017467248908,"run_count":1},"mail_phishing":{"rows":1050,"kind":"text","mean":0.9942857142857143,"runs":[0.9942857142857143],"min":0.9942857142857143,"max":0.9942857142857143,"run_count":1},"mail_spam":{"rows":854,"kind":"text","mean":0.9894613583138173,"runs":[0.9894613583138173],"min":0.9894613583138173,"max":0.9894613583138173,"run_count":1},"massive_en-US":{"rows":2970,"kind":"text","mean":0.8919191919191919,"runs":[0.8919191919191919],"min":0.8919191919191919,"max":0.8919191919191919,"run_count":1},"massive_zh-CN":{"rows":2921,"kind":"text","mean":0.8671687778158165,"runs":[0.8671687778158165],"min":0.8671687778158165,"max":0.8671687778158165,"run_count":1},"sharc":{"rows":8276,"kind":"text","mean":0.778757854035766,"runs":[0.778757854035766],"min":0.778757854035766,"max":0.778757854035766,"run_count":1},"sms_spam":{"rows":794,"kind":"text","mean":0.9937027707808564,"runs":[0.9937027707808564],"min":0.9937027707808564,"max":0.9937027707808564,"run_count":1},"typed_decisions":{"rows":2000,"kind":"text","mean":0.7865,"runs":[0.7865],"min":0.7865,"max":0.7865,"run_count":1},"wikiqa":{"rows":6160,"kind":"text","mean":0.9647727272727272,"runs":[0.9647727272727272],"min":0.9647727272727272,"max":0.9647727272727272,"run_count":1},"xnli_en":{"rows":5009,"kind":"text","mean":0.9147534438011579,"runs":[0.9147534438011579],"min":0.9147534438011579,"max":0.9147534438011579,"run_count":1},"xnli_zh":{"rows":5009,"kind":"text","mean":0.8434817328808145,"runs":[0.8434817328808145],"min":0.8434817328808145,"max":0.8434817328808145,"run_count":1},"aokvqa":{"rows":1138,"kind":"image","mean":0.8919156414762741,"runs":[0.8919156414762741],"min":0.8919156414762741,"max":0.8919156414762741,"run_count":1},"clevr_attribute":{"rows":53734,"kind":"image","mean":0.9991811515986154,"runs":[0.9991811515986154],"min":0.9991811515986154,"max":0.9991811515986154,"run_count":1},"clevr_count":{"rows":35422,"kind":"image","mean":0.9773869346733668,"runs":[0.9773869346733668],"min":0.9773869346733668,"max":0.9773869346733668,"run_count":1},"clevr_exist":{"rows":20196,"kind":"image","mean":0.9973262032085561,"runs":[0.9973262032085561],"min":0.9973262032085561,"max":0.9973262032085561,"run_count":1},"scienceqa":{"rows":2017,"kind":"image","mean":0.9796727813584531,"runs":[0.9796727813584531],"min":0.9796727813584531,"max":0.9796727813584531,"run_count":1},"screenqa_choice":{"rows":848,"kind":"image","mean":0.30778301886792453,"runs":[0.30778301886792453],"min":0.30778301886792453,"max":0.30778301886792453,"run_count":1},"screenqa_noul":{"rows":2148,"kind":"image","mean":0.74487895716946,"runs":[0.74487895716946],"min":0.74487895716946,"max":0.74487895716946,"run_count":1},"vqav2_yesno":{"rows":8102,"kind":"image","mean":0.8920019748210318,"runs":[0.8920019748210318],"min":0.8920019748210318,"max":0.8920019748210318,"run_count":1}}},{"id":"dohnuts_release","name":"Dohnuts 0.8B (released)","family":"Dohnuts","base":"Qwen3.5-0.8B","note":"Released Dohnuts checkpoint evaluated as a historical reference. This is a separately trained checkpoint, not the Jolt retrain.","run_count":1,"runs":[{"seed":"single","scope":{}}],"scores":{"text":{"macro":0.7864329665207999,"weighted":0.792542444972176,"macro_sd":null,"macro_min":0.7864329665207999,"macro_max":0.7864329665207999,"weighted_sd":null,"rows":56426,"datasets":18,"run_count":1,"weighted_min":0.792542444972176,"weighted_max":0.792542444972176},"image":{"macro":0.7703751188262005,"weighted":0.9138141660936047,"macro_sd":null,"macro_min":0.7703751188262005,"macro_max":0.7703751188262005,"weighted_sd":null,"rows":123605,"datasets":8,"run_count":1,"weighted_min":0.9138141660936047,"weighted_max":0.9138141660936047},"all":{"macro":0.781492090307077,"weighted":0.8758047225200104,"macro_sd":null,"macro_min":0.781492090307077,"macro_max":0.781492090307077,"weighted_sd":null,"rows":180031,"datasets":26,"run_count":1,"weighted_min":0.8758047225200104,"weighted_max":0.8758047225200104}},"by_dataset":{"sharc":{"rows":8276,"kind":"text","mean":0.6684388593523442,"runs":[0.6684388593523442],"min":0.6684388593523442,"max":0.6684388593523442,"run_count":1},"contract_nli":{"rows":1173,"kind":"text","mean":0.815004262574595,"runs":[0.815004262574595],"min":0.815004262574595,"max":0.815004262574595,"run_count":1},"wikiqa":{"rows":6160,"kind":"text","mean":0.9594155844155844,"runs":[0.9594155844155844],"min":0.9594155844155844,"max":0.9594155844155844,"run_count":1},"esci_us":{"rows":1145,"kind":"text","mean":0.5048034934497817,"runs":[0.5048034934497817],"min":0.5048034934497817,"max":0.5048034934497817,"run_count":1},"esci_es":{"rows":1482,"kind":"text","mean":0.5614035087719298,"runs":[0.5614035087719298],"min":0.5614035087719298,"max":0.5614035087719298,"run_count":1},"esci_jp":{"rows":1633,"kind":"text","mean":0.6007348438456828,"runs":[0.6007348438456828],"min":0.6007348438456828,"max":0.6007348438456828,"run_count":1},"mail_spam":{"rows":854,"kind":"text","mean":0.9871194379391101,"runs":[0.9871194379391101],"min":0.9871194379391101,"max":0.9871194379391101,"run_count":1},"mail_phishing":{"rows":1050,"kind":"text","mean":0.9923809523809524,"runs":[0.9923809523809524],"min":0.9923809523809524,"max":0.9923809523809524,"run_count":1},"sms_spam":{"rows":794,"kind":"text","mean":0.9937027707808564,"runs":[0.9937027707808564],"min":0.9937027707808564,"max":0.9937027707808564,"run_count":1},"massive_en-US":{"rows":2970,"kind":"text","mean":0.7915824915824916,"runs":[0.7915824915824916],"min":0.7915824915824916,"max":0.7915824915824916,"run_count":1},"massive_zh-CN":{"rows":2921,"kind":"text","mean":0.7548784662786717,"runs":[0.7548784662786717],"min":0.7548784662786717,"max":0.7548784662786717,"run_count":1},"boolq":{"rows":3270,"kind":"text","mean":0.8290519877675842,"runs":[0.8290519877675842],"min":0.8290519877675842,"max":0.8290519877675842,"run_count":1},"ag_news":{"rows":7600,"kind":"text","mean":0.895,"runs":[0.895],"min":0.895,"max":0.895,"run_count":1},"emotion":{"rows":2000,"kind":"text","mean":0.764,"runs":[0.764],"min":0.764,"max":0.764,"run_count":1},"banking77":{"rows":3080,"kind":"text","mean":0.7295454545454545,"runs":[0.7295454545454545],"min":0.7295454545454545,"max":0.7295454545454545,"run_count":1},"xnli_en":{"rows":5009,"kind":"text","mean":0.8384907167099221,"runs":[0.8384907167099221],"min":0.8384907167099221,"max":0.8384907167099221,"run_count":1},"xnli_zh":{"rows":5009,"kind":"text","mean":0.755240566979437,"runs":[0.755240566979437],"min":0.755240566979437,"max":0.755240566979437,"run_count":1},"typed_decisions":{"rows":2000,"kind":"text","mean":0.715,"runs":[0.715],"min":0.715,"max":0.715,"run_count":1},"aokvqa":{"rows":1138,"kind":"image","mean":0.7838312829525483,"runs":[0.7838312829525483],"min":0.7838312829525483,"max":0.7838312829525483,"run_count":1},"scienceqa":{"rows":2017,"kind":"image","mean":0.8448190381755082,"runs":[0.8448190381755082],"min":0.8448190381755082,"max":0.8448190381755082,"run_count":1},"vqav2_yesno":{"rows":8102,"kind":"image","mean":0.8090594914835842,"runs":[0.8090594914835842],"min":0.8090594914835842,"max":0.8090594914835842,"run_count":1},"clevr_exist":{"rows":20196,"kind":"image","mean":0.9740047534165182,"runs":[0.9740047534165182],"min":0.9740047534165182,"max":0.9740047534165182,"run_count":1},"clevr_attribute":{"rows":53734,"kind":"image","mean":0.9834369300629024,"runs":[0.9834369300629024],"min":0.9834369300629024,"max":0.9834369300629024,"run_count":1},"clevr_count":{"rows":35422,"kind":"image","mean":0.8347919372141607,"runs":[0.8347919372141607],"min":0.8347919372141607,"max":0.8347919372141607,"run_count":1},"screenqa_choice":{"rows":848,"kind":"image","mean":0.22169811320754718,"runs":[0.22169811320754718],"min":0.22169811320754718,"max":0.22169811320754718,"run_count":1},"screenqa_noul":{"rows":2148,"kind":"image","mean":0.7113594040968343,"runs":[0.7113594040968343],"min":0.7113594040968343,"max":0.7113594040968343,"run_count":1}}},{"id":"laya_multi","name":"Laya Multilingual","family":"Laya","base":"Laya multilingual","note":"Single checkpoint. Its complete training-data manifest is not public, so test-set exposure cannot be fully audited.","run_count":1,"runs":[{"seed":"single","scope":{}}],"scores":{"text":{"macro":0.5624994432347215,"weighted":0.6175699145783858,"macro_sd":null,"macro_min":0.5624994432347215,"macro_max":0.5624994432347215,"weighted_sd":null,"rows":56426,"datasets":18,"run_count":1,"weighted_min":0.6175699145783858,"weighted_max":0.6175699145783858},"image":null,"all":null},"by_dataset":{"sharc":{"rows":8276,"kind":"text","mean":0.3284195263412276,"runs":[0.3284195263412276],"min":0.3284195263412276,"max":0.3284195263412276,"run_count":1},"contract_nli":{"rows":1173,"kind":"text","mean":0.22762148337595908,"runs":[0.22762148337595908],"min":0.22762148337595908,"max":0.22762148337595908,"run_count":1},"wikiqa":{"rows":6160,"kind":"text","mean":0.9412337662337662,"runs":[0.9412337662337662],"min":0.9412337662337662,"max":0.9412337662337662,"run_count":1},"esci_us":{"rows":1145,"kind":"text","mean":0.25589519650655024,"runs":[0.25589519650655024],"min":0.25589519650655024,"max":0.25589519650655024,"run_count":1},"esci_es":{"rows":1482,"kind":"text","mean":0.29284750337381915,"runs":[0.29284750337381915],"min":0.29284750337381915,"max":0.29284750337381915,"run_count":1},"esci_jp":{"rows":1633,"kind":"text","mean":0.288426209430496,"runs":[0.288426209430496],"min":0.288426209430496,"max":0.288426209430496,"run_count":1},"mail_spam":{"rows":854,"kind":"text","mean":0.9789227166276346,"runs":[0.9789227166276346],"min":0.9789227166276346,"max":0.9789227166276346,"run_count":1},"mail_phishing":{"rows":1050,"kind":"text","mean":0.9047619047619048,"runs":[0.9047619047619048],"min":0.9047619047619048,"max":0.9047619047619048,"run_count":1},"sms_spam":{"rows":794,"kind":"text","mean":0.5617128463476071,"runs":[0.5617128463476071],"min":0.5617128463476071,"max":0.5617128463476071,"run_count":1},"massive_en-US":{"rows":2970,"kind":"text","mean":0.48114478114478115,"runs":[0.48114478114478115],"min":0.48114478114478115,"max":0.48114478114478115,"run_count":1},"massive_zh-CN":{"rows":2921,"kind":"text","mean":0.39438548442314275,"runs":[0.39438548442314275],"min":0.39438548442314275,"max":0.39438548442314275,"run_count":1},"boolq":{"rows":3270,"kind":"text","mean":0.6859327217125383,"runs":[0.6859327217125383],"min":0.6859327217125383,"max":0.6859327217125383,"run_count":1},"ag_news":{"rows":7600,"kind":"text","mean":0.908157894736842,"runs":[0.908157894736842],"min":0.908157894736842,"max":0.908157894736842,"run_count":1},"emotion":{"rows":2000,"kind":"text","mean":0.5195,"runs":[0.5195],"min":0.5195,"max":0.5195,"run_count":1},"banking77":{"rows":3080,"kind":"text","mean":0.45811688311688314,"runs":[0.45811688311688314],"min":0.45811688311688314,"max":0.45811688311688314,"run_count":1},"xnli_en":{"rows":5009,"kind":"text","mean":0.8253144340187663,"runs":[0.8253144340187663],"min":0.8253144340187663,"max":0.8253144340187663,"run_count":1},"xnli_zh":{"rows":5009,"kind":"text","mean":0.7240966260730685,"runs":[0.7240966260730685],"min":0.7240966260730685,"max":0.7240966260730685,"run_count":1},"typed_decisions":{"rows":2000,"kind":"text","mean":0.3485,"runs":[0.3485],"min":0.3485,"max":0.3485,"run_count":1},"aokvqa":{"rows":1138,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"scienceqa":{"rows":2017,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"vqav2_yesno":{"rows":8102,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"clevr_exist":{"rows":20196,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"clevr_attribute":{"rows":53734,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"clevr_count":{"rows":35422,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"screenqa_choice":{"rows":848,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"screenqa_noul":{"rows":2148,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0}}},{"id":"laya_vision","name":"Laya Vision","family":"Laya","base":"Laya Vision","note":"Single checkpoint. Text results are out-of-domain diagnostics. Published fine-tuning sources overlap A-OKVQA and VQAv2 test-source splits.","run_count":1,"runs":[{"seed":"single","scope":{}}],"scores":{"text":{"macro":0.419315890304322,"weighted":0.3897848509552334,"macro_sd":null,"macro_min":0.419315890304322,"macro_max":0.419315890304322,"weighted_sd":null,"rows":56426,"datasets":18,"run_count":1,"weighted_min":0.3897848509552334,"weighted_max":0.3897848509552334},"image":{"macro":0.515280907467984,"weighted":0.45876784919703895,"macro_sd":null,"macro_min":0.515280907467984,"macro_max":0.515280907467984,"weighted_sd":null,"rows":123605,"datasets":8,"run_count":1,"weighted_min":0.45876784919703895,"weighted_max":0.45876784919703895},"all":{"macro":0.44884358789314105,"weighted":0.4371469358055002,"macro_sd":null,"macro_min":0.44884358789314105,"macro_max":0.44884358789314105,"weighted_sd":null,"rows":180031,"datasets":26,"run_count":1,"weighted_min":0.4371469358055002,"weighted_max":0.4371469358055002}},"by_dataset":{"sharc":{"rows":8276,"kind":"text","mean":0.347027549540841,"runs":[0.347027549540841],"min":0.347027549540841,"max":0.347027549540841,"run_count":1},"contract_nli":{"rows":1173,"kind":"text","mean":0.4373401534526854,"runs":[0.4373401534526854],"min":0.4373401534526854,"max":0.4373401534526854,"run_count":1},"wikiqa":{"rows":6160,"kind":"text","mean":0.04756493506493507,"runs":[0.04756493506493507],"min":0.04756493506493507,"max":0.04756493506493507,"run_count":1},"esci_us":{"rows":1145,"kind":"text","mean":0.3720524017467249,"runs":[0.3720524017467249],"min":0.3720524017467249,"max":0.3720524017467249,"run_count":1},"esci_es":{"rows":1482,"kind":"text","mean":0.36774628879892035,"runs":[0.36774628879892035],"min":0.36774628879892035,"max":0.36774628879892035,"run_count":1},"esci_jp":{"rows":1633,"kind":"text","mean":0.3998775260257195,"runs":[0.3998775260257195],"min":0.3998775260257195,"max":0.3998775260257195,"run_count":1},"mail_spam":{"rows":854,"kind":"text","mean":0.4613583138173302,"runs":[0.4613583138173302],"min":0.4613583138173302,"max":0.4613583138173302,"run_count":1},"mail_phishing":{"rows":1050,"kind":"text","mean":0.6285714285714286,"runs":[0.6285714285714286],"min":0.6285714285714286,"max":0.6285714285714286,"run_count":1},"sms_spam":{"rows":794,"kind":"text","mean":0.8161209068010076,"runs":[0.8161209068010076],"min":0.8161209068010076,"max":0.8161209068010076,"run_count":1},"massive_en-US":{"rows":2970,"kind":"text","mean":0.3053872053872054,"runs":[0.3053872053872054],"min":0.3053872053872054,"max":0.3053872053872054,"run_count":1},"massive_zh-CN":{"rows":2921,"kind":"text","mean":0.1359123587812393,"runs":[0.1359123587812393],"min":0.1359123587812393,"max":0.1359123587812393,"run_count":1},"boolq":{"rows":3270,"kind":"text","mean":0.6446483180428134,"runs":[0.6446483180428134],"min":0.6446483180428134,"max":0.6446483180428134,"run_count":1},"ag_news":{"rows":7600,"kind":"text","mean":0.7294736842105263,"runs":[0.7294736842105263],"min":0.7294736842105263,"max":0.7294736842105263,"run_count":1},"emotion":{"rows":2000,"kind":"text","mean":0.487,"runs":[0.487],"min":0.487,"max":0.487,"run_count":1},"banking77":{"rows":3080,"kind":"text","mean":0.2948051948051948,"runs":[0.2948051948051948],"min":0.2948051948051948,"max":0.2948051948051948,"run_count":1},"xnli_en":{"rows":5009,"kind":"text","mean":0.3333998802156119,"runs":[0.3333998802156119],"min":0.3333998802156119,"max":0.3333998802156119,"run_count":1},"xnli_zh":{"rows":5009,"kind":"text","mean":0.3333998802156119,"runs":[0.3333998802156119],"min":0.3333998802156119,"max":0.3333998802156119,"run_count":1},"typed_decisions":{"rows":2000,"kind":"text","mean":0.406,"runs":[0.406],"min":0.406,"max":0.406,"run_count":1},"aokvqa":{"rows":1138,"kind":"image","mean":0.6221441124780316,"runs":[0.6221441124780316],"min":0.6221441124780316,"max":0.6221441124780316,"run_count":1},"scienceqa":{"rows":2017,"kind":"image","mean":0.8869608329201785,"runs":[0.8869608329201785],"min":0.8869608329201785,"max":0.8869608329201785,"run_count":1},"vqav2_yesno":{"rows":8102,"kind":"image","mean":0.7696864971611947,"runs":[0.7696864971611947],"min":0.7696864971611947,"max":0.7696864971611947,"run_count":1},"clevr_exist":{"rows":20196,"kind":"image","mean":0.5732323232323232,"runs":[0.5732323232323232],"min":0.5732323232323232,"max":0.5732323232323232,"run_count":1},"clevr_attribute":{"rows":53734,"kind":"image","mean":0.5970893661368966,"runs":[0.5970893661368966],"min":0.5970893661368966,"max":0.5970893661368966,"run_count":1},"clevr_count":{"rows":35422,"kind":"image","mean":0.08940771272090792,"runs":[0.08940771272090792],"min":0.08940771272090792,"max":0.08940771272090792,"run_count":1},"screenqa_choice":{"rows":848,"kind":"image","mean":0.08372641509433962,"runs":[0.08372641509433962],"min":0.08372641509433962,"max":0.08372641509433962,"run_count":1},"screenqa_noul":{"rows":2148,"kind":"image","mean":0.5,"runs":[0.5],"min":0.5,"max":0.5,"run_count":1}}},{"id":"kev_08","name":"Kev 0.8B","family":"Kev","base":"Kev 0.8B","note":"Single evaluation run. Kev training-source overlap is documented for AG News, Banking77, and BoolQ; XNLI is related to its MultiNLI source.","run_count":1,"runs":[{"seed":"single","scope":{}}],"scores":{"text":{"macro":0.6217632287540448,"weighted":0.6541310743274377,"macro_sd":null,"macro_min":0.6217632287540448,"macro_max":0.6217632287540448,"weighted_sd":null,"rows":56426,"datasets":18,"run_count":1,"weighted_min":0.6541310743274377,"weighted_max":0.6541310743274377},"image":null,"all":null},"by_dataset":{"sharc":{"rows":8276,"kind":"text","mean":0.3878685355244079,"runs":[0.3878685355244079],"min":0.3878685355244079,"max":0.3878685355244079,"run_count":1},"contract_nli":{"rows":1173,"kind":"text","mean":0.5174765558397272,"runs":[0.5174765558397272],"min":0.5174765558397272,"max":0.5174765558397272,"run_count":1},"wikiqa":{"rows":6160,"kind":"text","mean":0.7055194805194805,"runs":[0.7055194805194805],"min":0.7055194805194805,"max":0.7055194805194805,"run_count":1},"esci_us":{"rows":1145,"kind":"text","mean":0.39039301310043667,"runs":[0.39039301310043667],"min":0.39039301310043667,"max":0.39039301310043667,"run_count":1},"esci_es":{"rows":1482,"kind":"text","mean":0.3974358974358974,"runs":[0.3974358974358974],"min":0.3974358974358974,"max":0.3974358974358974,"run_count":1},"esci_jp":{"rows":1633,"kind":"text","mean":0.5058175137783221,"runs":[0.5058175137783221],"min":0.5058175137783221,"max":0.5058175137783221,"run_count":1},"mail_spam":{"rows":854,"kind":"text","mean":0.5971896955503513,"runs":[0.5971896955503513],"min":0.5971896955503513,"max":0.5971896955503513,"run_count":1},"mail_phishing":{"rows":1050,"kind":"text","mean":0.6857142857142857,"runs":[0.6857142857142857],"min":0.6857142857142857,"max":0.6857142857142857,"run_count":1},"sms_spam":{"rows":794,"kind":"text","mean":0.8073047858942065,"runs":[0.8073047858942065],"min":0.8073047858942065,"max":0.8073047858942065,"run_count":1},"massive_en-US":{"rows":2970,"kind":"text","mean":0.6191919191919192,"runs":[0.6191919191919192],"min":0.6191919191919192,"max":0.6191919191919192,"run_count":1},"massive_zh-CN":{"rows":2921,"kind":"text","mean":0.5994522423827456,"runs":[0.5994522423827456],"min":0.5994522423827456,"max":0.5994522423827456,"run_count":1},"boolq":{"rows":3270,"kind":"text","mean":0.8003058103975536,"runs":[0.8003058103975536],"min":0.8003058103975536,"max":0.8003058103975536,"run_count":1},"ag_news":{"rows":7600,"kind":"text","mean":0.890921052631579,"runs":[0.890921052631579],"min":0.890921052631579,"max":0.890921052631579,"run_count":1},"emotion":{"rows":2000,"kind":"text","mean":0.499,"runs":[0.499],"min":0.499,"max":0.499,"run_count":1},"banking77":{"rows":3080,"kind":"text","mean":0.8064935064935065,"runs":[0.8064935064935065],"min":0.8064935064935065,"max":0.8064935064935065,"run_count":1},"xnli_en":{"rows":5009,"kind":"text","mean":0.81493312038331,"runs":[0.81493312038331],"min":0.81493312038331,"max":0.81493312038331,"run_count":1},"xnli_zh":{"rows":5009,"kind":"text","mean":0.7107207027350768,"runs":[0.7107207027350768],"min":0.7107207027350768,"max":0.7107207027350768,"run_count":1},"typed_decisions":{"rows":2000,"kind":"text","mean":0.456,"runs":[0.456],"min":0.456,"max":0.456,"run_count":1},"aokvqa":{"rows":1138,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"scienceqa":{"rows":2017,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"vqav2_yesno":{"rows":8102,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"clevr_exist":{"rows":20196,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"clevr_attribute":{"rows":53734,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"clevr_count":{"rows":35422,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"screenqa_choice":{"rows":848,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"screenqa_noul":{"rows":2148,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0}}},{"id":"kev_4","name":"Kev 4B","family":"Kev","base":"Kev 4B","note":"Single evaluation run. Kev training-source overlap is documented for AG News, Banking77, and BoolQ; XNLI is related to its MultiNLI source.","run_count":1,"runs":[{"seed":"single","scope":{}}],"scores":{"text":{"macro":0.7455476250486927,"weighted":0.7756885123879063,"macro_sd":null,"macro_min":0.7455476250486927,"macro_max":0.7455476250486927,"weighted_sd":null,"rows":56426,"datasets":18,"run_count":1,"weighted_min":0.7756885123879063,"weighted_max":0.7756885123879063},"image":null,"all":null},"by_dataset":{"sharc":{"rows":8276,"kind":"text","mean":0.6468100531657806,"runs":[0.6468100531657806],"min":0.6468100531657806,"max":0.6468100531657806,"run_count":1},"contract_nli":{"rows":1173,"kind":"text","mean":0.7749360613810742,"runs":[0.7749360613810742],"min":0.7749360613810742,"max":0.7749360613810742,"run_count":1},"wikiqa":{"rows":6160,"kind":"text","mean":0.9412337662337662,"runs":[0.9412337662337662],"min":0.9412337662337662,"max":0.9412337662337662,"run_count":1},"esci_us":{"rows":1145,"kind":"text","mean":0.4794759825327511,"runs":[0.4794759825327511],"min":0.4794759825327511,"max":0.4794759825327511,"run_count":1},"esci_es":{"rows":1482,"kind":"text","mean":0.47233468286099867,"runs":[0.47233468286099867],"min":0.47233468286099867,"max":0.47233468286099867,"run_count":1},"esci_jp":{"rows":1633,"kind":"text","mean":0.5762400489895897,"runs":[0.5762400489895897],"min":0.5762400489895897,"max":0.5762400489895897,"run_count":1},"mail_spam":{"rows":854,"kind":"text","mean":0.9543325526932084,"runs":[0.9543325526932084],"min":0.9543325526932084,"max":0.9543325526932084,"run_count":1},"mail_phishing":{"rows":1050,"kind":"text","mean":0.700952380952381,"runs":[0.700952380952381],"min":0.700952380952381,"max":0.700952380952381,"run_count":1},"sms_spam":{"rows":794,"kind":"text","mean":0.9534005037783375,"runs":[0.9534005037783375],"min":0.9534005037783375,"max":0.9534005037783375,"run_count":1},"massive_en-US":{"rows":2970,"kind":"text","mean":0.7047138047138047,"runs":[0.7047138047138047],"min":0.7047138047138047,"max":0.7047138047138047,"run_count":1},"massive_zh-CN":{"rows":2921,"kind":"text","mean":0.6881205066757959,"runs":[0.6881205066757959],"min":0.6881205066757959,"max":0.6881205066757959,"run_count":1},"boolq":{"rows":3270,"kind":"text","mean":0.8990825688073395,"runs":[0.8990825688073395],"min":0.8990825688073395,"max":0.8990825688073395,"run_count":1},"ag_news":{"rows":7600,"kind":"text","mean":0.9035526315789474,"runs":[0.9035526315789474],"min":0.9035526315789474,"max":0.9035526315789474,"run_count":1},"emotion":{"rows":2000,"kind":"text","mean":0.5675,"runs":[0.5675],"min":0.5675,"max":0.5675,"run_count":1},"banking77":{"rows":3080,"kind":"text","mean":0.8555194805194806,"runs":[0.8555194805194806],"min":0.8555194805194806,"max":0.8555194805194806,"run_count":1},"xnli_en":{"rows":5009,"kind":"text","mean":0.8818127370732681,"runs":[0.8818127370732681],"min":0.8818127370732681,"max":0.8818127370732681,"run_count":1},"xnli_zh":{"rows":5009,"kind":"text","mean":0.7558394889199441,"runs":[0.7558394889199441],"min":0.7558394889199441,"max":0.7558394889199441,"run_count":1},"typed_decisions":{"rows":2000,"kind":"text","mean":0.664,"runs":[0.664],"min":0.664,"max":0.664,"run_count":1},"aokvqa":{"rows":1138,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"scienceqa":{"rows":2017,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"vqav2_yesno":{"rows":8102,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"clevr_exist":{"rows":20196,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"clevr_attribute":{"rows":53734,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"clevr_count":{"rows":35422,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"screenqa_choice":{"rows":848,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"screenqa_noul":{"rows":2148,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0}}},{"id":"kev_9","name":"Kev 9B","family":"Kev","base":"Kev 9B","note":"Single evaluation run. Kev training-source overlap is documented for AG News, Banking77, and BoolQ; XNLI is related to its MultiNLI source.","run_count":1,"runs":[{"seed":"single","scope":{}}],"scores":{"text":{"macro":0.7684666731466621,"weighted":0.78277744302272,"macro_sd":null,"macro_min":0.7684666731466621,"macro_max":0.7684666731466621,"weighted_sd":null,"rows":56426,"datasets":18,"run_count":1,"weighted_min":0.78277744302272,"weighted_max":0.78277744302272},"image":null,"all":null},"by_dataset":{"sharc":{"rows":8276,"kind":"text","mean":0.6210729821169647,"runs":[0.6210729821169647],"min":0.6210729821169647,"max":0.6210729821169647,"run_count":1},"contract_nli":{"rows":1173,"kind":"text","mean":0.7902813299232737,"runs":[0.7902813299232737],"min":0.7902813299232737,"max":0.7902813299232737,"run_count":1},"wikiqa":{"rows":6160,"kind":"text","mean":0.9157467532467533,"runs":[0.9157467532467533],"min":0.9157467532467533,"max":0.9157467532467533,"run_count":1},"esci_us":{"rows":1145,"kind":"text","mean":0.5056768558951965,"runs":[0.5056768558951965],"min":0.5056768558951965,"max":0.5056768558951965,"run_count":1},"esci_es":{"rows":1482,"kind":"text","mean":0.5641025641025641,"runs":[0.5641025641025641],"min":0.5641025641025641,"max":0.5641025641025641,"run_count":1},"esci_jp":{"rows":1633,"kind":"text","mean":0.6117575015309247,"runs":[0.6117575015309247],"min":0.6117575015309247,"max":0.6117575015309247,"run_count":1},"mail_spam":{"rows":854,"kind":"text","mean":0.9508196721311475,"runs":[0.9508196721311475],"min":0.9508196721311475,"max":0.9508196721311475,"run_count":1},"mail_phishing":{"rows":1050,"kind":"text","mean":0.8866666666666667,"runs":[0.8866666666666667],"min":0.8866666666666667,"max":0.8866666666666667,"run_count":1},"sms_spam":{"rows":794,"kind":"text","mean":0.9672544080604534,"runs":[0.9672544080604534],"min":0.9672544080604534,"max":0.9672544080604534,"run_count":1},"massive_en-US":{"rows":2970,"kind":"text","mean":0.7252525252525253,"runs":[0.7252525252525253],"min":0.7252525252525253,"max":0.7252525252525253,"run_count":1},"massive_zh-CN":{"rows":2921,"kind":"text","mean":0.7038685381718589,"runs":[0.7038685381718589],"min":0.7038685381718589,"max":0.7038685381718589,"run_count":1},"boolq":{"rows":3270,"kind":"text","mean":0.9045871559633027,"runs":[0.9045871559633027],"min":0.9045871559633027,"max":0.9045871559633027,"run_count":1},"ag_news":{"rows":7600,"kind":"text","mean":0.9025,"runs":[0.9025],"min":0.9025,"max":0.9025,"run_count":1},"emotion":{"rows":2000,"kind":"text","mean":0.574,"runs":[0.574],"min":0.574,"max":0.574,"run_count":1},"banking77":{"rows":3080,"kind":"text","mean":0.865909090909091,"runs":[0.865909090909091],"min":0.865909090909091,"max":0.865909090909091,"run_count":1},"xnli_en":{"rows":5009,"kind":"text","mean":0.8931922539429028,"runs":[0.8931922539429028],"min":0.8931922539429028,"max":0.8931922539429028,"run_count":1},"xnli_zh":{"rows":5009,"kind":"text","mean":0.7712118187262926,"runs":[0.7712118187262926],"min":0.7712118187262926,"max":0.7712118187262926,"run_count":1},"typed_decisions":{"rows":2000,"kind":"text","mean":0.6785,"runs":[0.6785],"min":0.6785,"max":0.6785,"run_count":1},"aokvqa":{"rows":1138,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"scienceqa":{"rows":2017,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"vqav2_yesno":{"rows":8102,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"clevr_exist":{"rows":20196,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"clevr_attribute":{"rows":53734,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"clevr_count":{"rows":35422,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"screenqa_choice":{"rows":848,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0},"screenqa_noul":{"rows":2148,"kind":"image","mean":null,"runs":[],"min":null,"max":null,"run_count":0}}}]}; | |
| const DIAG={"screens":[{"id":"B0","name":"Dohnuts 0.8B (retrained)","change":"Frozen Dohnuts RLCD baseline recipe","s400":68.0,"s800":70.76,"fullDev":77.93},{"id":"L1","name":"Jolt L1","change":"Soft-target cross-entropy only","s400":74.58,"s800":76.23,"fullDev":80.57},{"id":"L2","name":"Jolt 0.8B L2","change":"Direct composite loss: CE, spherical score, ordinal RPS","s400":74.31,"s800":76.11,"fullDev":80.78},{"id":"L3","name":"Jolt L3","change":"Cross-entropy plus pathwise noisy reward","s400":74.35,"s800":76.38,"fullDev":80.52},{"id":"D1","name":"Jolt D1","change":"Sample without replacement","s400":69.65,"s800":73.18,"fullDev":null},{"id":"D2","name":"Jolt D2","change":"Expand eligible training pool, remove 6,000-row cap","s400":64.99,"s800":71.42,"fullDev":null},{"id":"H1","name":"Jolt H1","change":"Halve backbone and scorer learning rates","s400":66.44,"s800":70.64,"fullDev":null},{"id":"R1","name":"Jolt R1","change":"LoRA rank and alpha 8/16 to 16/32","s400":66.84,"s800":71.47,"fullDev":null},{"id":"A1","name":"Jolt A1","change":"Residual mixture of last four decoder layers","s400":68.07,"s800":70.24,"fullDev":null},{"id":"A2","name":"Jolt A2","change":"Candidate-marker attention","s400":69.08,"s800":69.77,"fullDev":null},{"id":"A3","name":"Jolt A3","change":"SiLU residual adapters in decoder blocks 16 to 23","s400":67.05,"s800":71.09,"fullDev":null},{"id":"A3G","name":"Jolt A3G","change":"A3 adapters with GELU activation","s400":68.48,"s800":72.37,"fullDev":72.52},{"id":"A4","name":"Jolt A4","change":"Shared GELU MLP candidate scorer","s400":70.12,"s800":71.63,"fullDev":null},{"id":"A5","name":"Jolt A5","change":"Learned SiLU/GELU mixture in last four feed-forward blocks","s400":68.43,"s800":71.57,"fullDev":null},{"id":"A5C","name":"Jolt A5C","change":"Unfused SiLU control matched to A5","s400":68.83,"s800":71.57,"fullDev":null},{"id":"P1","name":"Jolt P1","change":"Move candidate descriptions before readout markers","s400":69.55,"s800":71.66,"fullDev":77.92},{"id":"B0P1","name":"Jolt B0P1","change":"B0 prompt layout on P1 eligible training pool","s400":69.55,"s800":71.0,"fullDev":78.12}],"kevDev":[{"model":"Kev 0.8B","suite":"decision-v7","rows":1264,"score":82.8},{"model":"Kev 0.8B","suite":"transfer-v4","rows":656,"score":64.8},{"model":"Kev 0.8B","suite":"transfer-v9","rows":1046,"score":58.5},{"model":"Kev 4B","suite":"decision-v7","rows":1264,"score":87.2},{"model":"Kev 4B","suite":"transfer-v4","rows":656,"score":81.6},{"model":"Kev 4B","suite":"transfer-v9","rows":1046,"score":77.2},{"model":"Kev 9B","suite":"decision-v7","rows":1264,"score":87.4},{"model":"Kev 9B","suite":"transfer-v4","rows":656,"score":81.9},{"model":"Kev 9B","suite":"transfer-v9","rows":1046,"score":78.0}],"secondary":[{"model":"Jolt 0.8B L2","suite":"Additional application suite collection","rows":27015,"score":70.8680362761,"macro":68.5419359598,"note":"111 suites; 100% valid responses. This was a secondary evaluation and did not select the checkpoint."}],"localJevbench":[{"id":"kev_9","name":"Kev 9B","correct":184,"total":231,"easy_correct":48,"easy_total":48,"standard_correct":70,"standard_total":72,"hard_correct":66,"hard_total":111,"rejected":0,"source":"Local evaluation; BF16, checkpoint temperature, Kev serving context."},{"id":"kev_4","name":"Kev 4B","correct":175,"total":231,"easy_correct":48,"easy_total":48,"standard_correct":67,"standard_total":72,"hard_correct":60,"hard_total":111,"rejected":0,"source":"Local evaluation; BF16, checkpoint temperature, Kev serving context."},{"id":"jolt_l2_2b","name":"Jolt 2B L2","correct":168,"total":231,"easy_correct":48,"easy_total":48,"standard_correct":66,"standard_total":72,"hard_correct":54,"hard_total":111,"rejected":0,"source":"Local reevaluation; seed 44, step 2,800."},{"id":"dohnuts_release","name":"Dohnuts 0.8B (released)","correct":152,"total":231,"easy_correct":48,"easy_total":48,"standard_correct":56,"standard_total":72,"hard_correct":48,"hard_total":111,"rejected":0,"source":"Saved prior local evaluation; 4,096-token limit."},{"id":"jolt_b0","name":"Dohnuts 0.8B (retrained)","correct":148,"total":231,"easy_correct":47,"easy_total":48,"standard_correct":57,"standard_total":72,"hard_correct":44,"hard_total":111,"rejected":0,"source":"Local reevaluation; seed 44, step 3,600."},{"id":"kev_08","name":"Kev 0.8B","correct":148,"total":231,"easy_correct":48,"easy_total":48,"standard_correct":59,"standard_total":72,"hard_correct":41,"hard_total":111,"rejected":0,"source":"Local evaluation; BF16, checkpoint temperature, Kev serving context."},{"id":"jolt_l2_08","name":"Jolt 0.8B L2","correct":147,"total":231,"easy_correct":48,"easy_total":48,"standard_correct":57,"standard_total":72,"hard_correct":42,"hard_total":111,"rejected":0,"source":"Saved prior local evaluation; seed 42, step 3,600."},{"id":"laya_vision","name":"Laya Vision","correct":111,"total":231,"easy_correct":43,"easy_total":48,"standard_correct":34,"standard_total":72,"hard_correct":34,"hard_total":111,"rejected":0,"source":"Saved prior local evaluation on text-only tasks; native 1,024-token truncation."},{"id":"laya_multi","name":"Laya Multilingual","correct":110,"total":231,"easy_correct":43,"easy_total":48,"standard_correct":30,"standard_total":72,"hard_correct":37,"hard_total":111,"rejected":0,"source":"Saved prior local evaluation; native 1,024-token truncation."},{"id":"jolt_l2_4b","name":"Jolt 4B L2","correct":181,"total":231,"easy_correct":48,"easy_total":48,"standard_correct":69,"standard_total":72,"hard_correct":64,"hard_total":111,"rejected":0,"source":"Local evaluation; BF16 inference, seed 42, step 3600; unit temperature, hard accuracy."},{"id":"jolt_l2_9b_bf16","name":"Jolt 9B L2","correct":182,"total":231,"easy_correct":48,"easy_total":48,"standard_correct":70,"standard_total":72,"hard_correct":64,"hard_total":111,"rejected":0,"source":"Local evaluation; BF16 inference, seed 42, step 3,600; unit temperature, hard accuracy."}]}; | |
| const MODEL_COLORS={jolt_b0:'#687b84',jolt_l2_08:'#087f78',jolt_l2_2b:'#3275b7',jolt_l2_4b:'#548235',dohnuts_release:'#a5774a',laya_multi:'#7653a6',laya_vision:'#ba6b98',kev_08:'#5d91c7',kev_4:'#3565a8',kev_9:'#233f70',jolt_l2_9b_bf16:'#b8860b'}; | |
| let state={scope:'text',metric:'macro',sort:{key:'rows',dir:'desc'},detailModels:['jolt_b0','jolt_l2_08','jolt_l2_2b']}; | |
| const $=id=>document.getElementById(id), pct=x=>x==null?'—':(x*100).toFixed(2)+'%', pp=x=>(x>=0?'+':'')+(x*100).toFixed(2)+' pp', num=x=>Number(x).toLocaleString('en-US'); | |
| function scopeDatasets(){return Object.entries(DATA.datasets).filter(([,v])=>state.scope==='all'||v.kind===state.scope).map(([k])=>k)} | |
| function supportedModels(){return DATA.models.filter(m=>m.scores[state.scope])} | |
| function selectedScore(m){return m.scores[state.scope][state.metric]} | |
| function scoreRange(m){let s=m.scores[state.scope];return state.metric==='weighted'?[s.weighted_min,s.weighted_max]:[s.macro_min,s.macro_max]} | |
| function sortModels(list){return list.slice().sort((a,b)=>(selectedScore(b)??-1)-(selectedScore(a)??-1))} | |
| function displayDataset(k){return k.replaceAll('_',' ').replaceAll('-',' ').replace(/\b[a-z]/g,x=>x.toUpperCase())} | |
| function esc(s){return String(s).replace(/[&<>"']/g,c=>({'&':'&','<':'<','>':'>','"':'"',"'":'''}[c]))} | |
| function setupDetailModels(list){let el=$('detailModels'),available=new Set(list.map(m=>m.id));state.detailModels=state.detailModels.filter(id=>available.has(id));if(!state.detailModels.length)state.detailModels=list.slice(0,Math.min(3,list.length)).map(m=>m.id);el.innerHTML=list.map(m=>`<label class="model-option"><input type="checkbox" value="${m.id}" ${state.detailModels.includes(m.id)?'checked':''}><i class="series-swatch" style="background:${MODEL_COLORS[m.id]}"></i><span>${esc(m.name)}</span></label>`).join('');$('detailSummary').textContent=`${state.detailModels.length} model${state.detailModels.length===1?'':'s'} selected`;el.querySelectorAll('input[type=checkbox]').forEach(input=>input.addEventListener('change',()=>{state.detailModels=Array.from(el.querySelectorAll('input[type=checkbox]:checked')).map(x=>x.value);$('detailSummary').textContent=`${state.detailModels.length} model${state.detailModels.length===1?'':'s'} selected`;renderDatasetDetail(sortModels(supportedModels()))}))} | |
| function render(){let list=sortModels(supportedModels());setupDetailModels(list);let scope=DATA.meta.scope_counts[state.scope];$('laneNote').textContent=`${scope.datasets} datasets · ${num(scope.rows)} examples. Unsupported models are omitted from this lane.`;let scopeLabel=state.scope==='text'?'text':state.scope==='image'?'image':'all-dataset';$('rankTitle').textContent=`${scopeLabel[0].toUpperCase()+scopeLabel.slice(1)} lane ranking`;$('rankSub').textContent=state.metric==='macro'?'Every dataset counts equally.':'Each example counts equally.';renderKpis(list,scope);renderRank(list);renderDatasetDetail(list);renderTable(list);renderInterpretation(list);renderExposure();renderLegend(list);} | |
| function renderKpis(list,scope){let best=list[0],next=list[1],score=best?selectedScore(best):null,lead=best&&next?score-selectedScore(next):null;$('bestValue').textContent=pct(score);$('bestName').textContent=best?best.name:'No model in view';$('leadValue').textContent=lead==null?'—':pp(lead);$('leadName').textContent=next?`ahead of ${next.name}`:'No second model in view';$('coverageValue').textContent=`${scope.datasets} / 26`; $('coverageName').textContent=`${num(scope.rows)} test examples in this lane`;$('runsValue').textContent=best?`${best.run_count} ${best.run_count===1?'run':'seeds'}`:'—';$('runsName').textContent=best?(best.run_count>1?'mean across independent Jolt seeds':'one evaluated checkpoint'):'—'} | |
| function renderRank(list){let svg=$('ranking'),W=820,rowH=54,top=28,left=245,right=70,bottom=28,H=top+list.length*rowH+bottom,plot=W-left-right;svg.setAttribute('viewBox',`0 0 ${W} ${H}`);svg.setAttribute('height',H);let out=[];for(let tick=0;tick<=100;tick+=20){let x=left+plot*tick/100;out.push(`<line x1="${x}" x2="${x}" y1="${top-5}" y2="${H-bottom}" stroke="#e5ebed"/><text x="${x}" y="${H-6}" text-anchor="middle" font-size="11" fill="#6b7a82">${tick}%</text>`)}list.forEach((m,i)=>{let y=top+i*rowH,score=selectedScore(m);if(score==null)return;let width=plot*score,range=scoreRange(m),hasRange=m.run_count>1&&range[0]!=null&&range[1]!=null;let x1=left+plot*range[0],x2=left+plot*range[1],name=m.name+(m.id==='laya_vision'&&state.scope==='text'?' · diagnostic':'');let badge=m.run_count>1?'3 seeds':'single run';let extra=m.id==='laya_vision'&&state.scope==='text'?' · out of domain':m.family==='Kev'?' · source overlap':m.id==='dohnuts_release'?' · historical reference':'';out.push(`<g class="rank-item" tabindex="0"><title>${esc(name)}: ${pct(score)}${hasRange?` (seed range ${pct(range[0])} to ${pct(range[1])})`:''}. ${esc(m.note)}</title><text class="rank-label" x="${left-12}" y="${y+18}" text-anchor="end" font-size="13" font-weight="650" fill="#223843">${esc(name)}</text><text x="${left-12}" y="${y+36}" text-anchor="end" font-size="10" fill="#71818a">${badge}${extra}</text><rect x="${left}" y="${y+5}" width="${plot}" height="25" rx="5" fill="#f0f3f3"/><rect x="${left}" y="${y+5}" width="${width}" height="25" rx="5" fill="${MODEL_COLORS[m.id]||'#657782'}" opacity=".9"/>${hasRange?`<line x1="${x1}" x2="${x2}" y1="${y+17}" y2="${y+17}" stroke="#142d38" stroke-width="2.5"/><line x1="${x1}" x2="${x1}" y1="${y+11}" y2="${y+23}" stroke="#142d38"/><line x1="${x2}" x2="${x2}" y1="${y+11}" y2="${y+23}" stroke="#142d38"/>`:''}<text x="${Math.min(left+width+7,W-55)}" y="${y+23}" font-size="13" font-weight="750" fill="#213742">${pct(score)}</text></g>`)});svg.innerHTML=out.join('');} | |
| function renderDatasetDetail(list){let chosen=state.detailModels.map(id=>DATA.models.find(m=>m.id===id)).filter(Boolean);let order=new Map(list.map((m,i)=>[m.id,i]));chosen.sort((a,b)=>(order.get(a.id)??Infinity)-(order.get(b.id)??Infinity));let svg=$('datasetChart');if(!chosen.length){svg.innerHTML='';$('detailLegend').innerHTML='';$('datasetCaption').textContent='Select at least one model to draw the comparison.';return}let ds=scopeDatasets().filter(d=>chosen.some(m=>m.by_dataset[d]?.mean!=null));ds.sort((a,b)=>chosen.reduce((s,m)=>s+(m.by_dataset[b]?.mean??0)-(m.by_dataset[a]?.mean??0),0));let W=Math.max(320,Math.floor(svg.parentElement?.clientWidth||980)),labelW=Math.min(190,Math.max(135,Math.floor(W*.34))),plotX=labelW+10,right=45,plot=W-plotX-right,headerH=32,modelRow=16,gap=7,top=5,bottom=18,groupH=headerH+chosen.length*modelRow+gap,H=top+ds.length*groupH+bottom;svg.setAttribute('viewBox',`0 0 ${W} ${H}`);svg.setAttribute('height',H);svg.setAttribute('width',W);let out=[];for(let tick=0;tick<=100;tick+=20){let x=plotX+plot*tick/100;out.push(`<line x1="${x}" x2="${x}" y1="${top-4}" y2="${H-bottom}" stroke="#e7edef"/><text x="${x}" y="${H-5}" text-anchor="${tick===0?'start':'middle'}" font-size="10" fill="#71818a">${tick}%</text>`)}ds.forEach((d,i)=>{let y=top+i*groupH;out.push(`<text x="0" y="${y+12}" font-size="11" font-weight="750" fill="#263d48">${esc(displayDataset(d))}</text><text x="0" y="${y+25}" font-size="9.5" fill="#71818a">${num(DATA.datasets[d].rows)} rows</text>`);chosen.forEach((m,j)=>{let a=m.by_dataset[d]?.mean;if(a==null)return;let yy=y+headerH+j*modelRow,w=a*plot,color=MODEL_COLORS[m.id]||'#657782';out.push(`<g><title>${esc(m.name)} · ${esc(displayDataset(d))}: ${pct(a)} (${num(DATA.datasets[d].rows)} examples; ${m.run_count>1?'mean of 3 seeds':'one run'})</title><text x="${plotX-8}" y="${yy+11}" text-anchor="end" font-size="9.5" fill="#435660">${esc(m.name)}</text><rect x="${plotX}" y="${yy+2}" width="${plot}" height="10" rx="4" fill="#f0f3f3"/><rect x="${plotX}" y="${yy+2}" width="${w}" height="10" rx="4" fill="${color}"/><text x="${Math.min(plotX+w+5,W-45)}" y="${yy+11}" font-size="10" font-weight="700" fill="#344850">${pct(a)}</text></g>`)});out.push(`<line x1="0" x2="${W}" y1="${y+groupH-gap/2}" y2="${y+groupH-gap/2}" stroke="#e4eaec"/>`)});svg.innerHTML=out.join('');$('detailLegend').innerHTML=chosen.map((m,i)=>`<span><i class="dot" style="background:${MODEL_COLORS[m.id]||'#657782'}"></i><b class="series-index">${i+1}.</b> ${esc(m.name)} <span class="badge ${m.run_count>1?'seed':'single'}">${m.run_count>1?'3 seeds':'single run'}</span></span>`).join('');$('datasetCaption').textContent=`Each dataset is one stacked block: its name and sample count appear above the bars. Bars run top to bottom in the numbered legend order. Jolt bars use the three-seed mean; other bars use the single evaluated checkpoint. Hover a bar for its exact score.`} | |
| function heat(v){if(v==null)return'#f1f3f3';let t=Math.max(0,Math.min(1,(v-.35)/.65));let r=Math.round(249+(224-249)*t),g=Math.round(226+(242-226)*t),b=Math.round(220+(235-220)*t);return`rgb(${r},${g},${b})`} | |
| function renderTable(list){let table=$('scoreTable'),ds=scopeDatasets().filter(d=>DATA.datasets[d]);let key=state.sort.key,dir=state.sort.dir;ds.sort((a,b)=>{let av=key==='rows'?DATA.datasets[a].rows:key==='dataset'?a:(DATA.models.find(m=>m.id===key)?.by_dataset[a]?.mean??-1),bv=key==='rows'?DATA.datasets[b].rows:key==='dataset'?b:(DATA.models.find(m=>m.id===key)?.by_dataset[b]?.mean??-1);return(av>bv?1:av<bv?-1:0)*(dir==='asc'?1:-1)});let header=`<thead><tr><th data-sort="dataset">Dataset ↕</th><th data-sort="rows">Rows ↕</th>${list.map(m=>`<th data-sort="${m.id}" title="${esc(m.note)}">${esc(m.name)}<br><span class="dataset-sub">${m.run_count>1?'mean · 3 seeds':'one run'}</span></th>`).join('')}</tr></thead>`;let body=ds.map(d=>`<tr><td>${esc(displayDataset(d))}<br><span class="dataset-sub">${DATA.datasets[d].kind}</span></td><td>${num(DATA.datasets[d].rows)}</td>${list.map(m=>{let val=m.by_dataset[d]?.mean;return`<td><span class="cell" style="background:${heat(val)}" title="${esc(m.name)}: ${pct(val)}">${pct(val)}</span></td>`}).join('')}</tr>`).join('');table.innerHTML=header+`<tbody>${body}</tbody>`;table.querySelectorAll('th[data-sort]').forEach(th=>th.onclick=()=>{let k=th.dataset.sort;state.sort={key:k,dir:state.sort.key===k&&state.sort.dir==='desc'?'asc':'desc'};renderTable(sortModels(supportedModels()))})} | |
| function renderInterpretation(list){ | |
| const scope=state.scope,joltL2=DATA.models.find(m=>m.id==='jolt_l2_08'),b0=DATA.models.find(m=>m.id==='jolt_b0'),j2=DATA.models.find(m=>m.id==='jolt_l2_2b'),j4=DATA.models.find(m=>m.id==='jolt_l2_4b'),j9=DATA.models.find(m=>m.id==='jolt_l2_9b'); | |
| const score=m=>m?.scores[scope]?.macro, gap=((score(j9)-score(j4))*100).toFixed(2), items=[]; | |
| const lane=scope==='text'?'text':scope==='image'?'image':'all-dataset'; | |
| items.push(`<div class="note"><h3>Jolt 9B tops the ${lane} lane</h3><p>Its dataset-macro score is ${pct(score(j9))}, ${gap} percentage points above Jolt 4B L2. Each 4B and 9B result is from one checkpoint; the 0.8B and 2B figures are three-seed means.</p></div>`); | |
| if(scope==='text'){ | |
| items.push(`<div class="note"><h3>The objective change beats the baseline</h3><p>Jolt 0.8B L2 is ${((score(joltL2)-score(b0))*100).toFixed(2)} percentage points above the retrained B0 baseline on text macro. Both are three-seed means from the same training campaign.</p></div>`); | |
| items.push(`<div class="note"><h3>Kev is a one-run reference</h3><p>Kev 9B scores ${pct(DATA.models.find(m=>m.id==='kev_9').scores.text.macro)} and has known task-source overlap on some rows. These scores do not have seed-level variation estimates.</p></div>`); | |
| }else if(scope==='image'){ | |
| items.push(`<div class="note"><h3>Image scores depend on source exposure</h3><p>Laya Vision and the released Dohnuts checkpoint have documented overlap with some image test sources. Their scores remain visible, with that exposure described below.</p></div>`); | |
| }else{ | |
| items.push(`<div class="note"><h3>Dataset size changes the row-weighted result</h3><p>Large image datasets carry more weight in the pooled score. Dataset macro gives small tasks the same influence as CLEVR and VQAv2.</p></div>`); | |
| } | |
| items.push(`<div class="note"><h3>The Jolt test rows stayed held out</h3><p>These Jolt checkpoints were fine-tuned on the training partition. Development selected checkpoints; separate calibration data fit temperatures; locked test rows were excluded from fine-tuning, selection, and calibration. Qwen pretraining exposure is outside this audit.</p></div>`); | |
| items.push(`<div class="note"><h3>Seed means are not ensembles</h3><p>The 0.8B and 2B Jolt scores average three checkpoints. Jolt 4B and 9B each have one evaluated checkpoint. None of these numbers averages predictions across checkpoints.</p></div>`); | |
| $('interpretation').innerHTML=items.join('') | |
| } | |
| function renderExposure(){let rows=[['Jolt models','clean campaign holdout','B0, 0.8B L2, 2B L2, 4B L2, and 9B L2 were fine-tuned on the Jolt train partition. Development selected checkpoints; separate calibration data fit temperatures; locked test rows were excluded from campaign training, selection, and calibration. Qwen pretraining exposure is outside this audit.'],['Dohnuts 0.8B (released)','historical checkpoint exposure','The original release reports checkpoint selection on an A-OKVQA validation split that supplies Dohnuts test rows. VQAv2 also has documented validation-set fine-tuning overlap. Treat these image scores as exposed reference measurements.'],['Laya Vision','known image-source overlap','Its card documents A-OKVQA, ScienceQA, and VQAv2 fine-tuning. A-OKVQA and VQAv2 rows in this benchmark come from validation sources used in that history. Text scores are outside the model’s intended evaluation scope.'],['Kev','known text-source overlap','AG News, Banking77, and BoolQ appear in Kev training sources. XNLI is related to Kev’s MultiNLI source. The shared score table shows the results but flags this limitation.'],['Laya Multilingual','training exposure not fully audited','The model card does not provide a complete training-row manifest, so overlap with the shared test cannot be ruled out.']];$('exposure').innerHTML=rows.map(([name,badge,text])=>`<div class="exposure-row"><strong>${esc(name)}</strong> <span class="badge ${badge.includes('clean')?'clean':'exposure'}">${esc(badge)}</span><p>${esc(text)}</p></div>`).join('')} | |
| function renderLegend(list){$('familyLegend').innerHTML=list.map(m=>`<span><i class="dot" style="background:${MODEL_COLORS[m.id]||'#657782'}"></i>${esc(m.name)}</span>`).join('')+'<span><i class="range-key"></i>seed min–max</span>'} | |
| function makeTable(el,heads,rows){el.innerHTML=`<thead><tr>${heads.map(h=>`<th>${h}</th>`).join('')}</tr></thead><tbody>${rows.map(row=>`<tr>${row.map(v=>`<td>${v}</td>`).join('')}</tr>`).join('')}</tbody>`} | |
| function renderLocalJevbench(){const balanced=m=>((m.easy_correct/m.easy_total)+(m.standard_correct/m.standard_total)+(m.hard_correct/m.hard_total))/3;let list=DIAG.localJevbench.slice().sort((a,b)=>balanced(b)-balanced(a));$('jevbenchChart').innerHTML=list.map(m=>{let score=balanced(m),color=MODEL_COLORS[m.id]||'#657782';return`<div class="diag-bar-row"><span title="${esc(m.source)}">${esc(m.name)}</span><div class="diag-track" title="Equal-weight tier score ${pct(score)}"><div class="diag-fill" style="width:${(score*100).toFixed(2)}%;background:${color}"></div></div><strong>${pct(score)}</strong></div>`}).join('');makeTable($('jevbenchTable'),['Model / evaluation','Tier-balanced headline','Pooled','Easy','Standard','Hard','Rejections / failures'],list.map(m=>[`${esc(m.name)}<br><span class="dataset-sub">${esc(m.source)}</span>`,pct(balanced(m)),`${m.correct}/${m.total} · ${pct(m.correct/m.total)}`,`${m.easy_correct}/${m.easy_total} · ${pct(m.easy_correct/m.easy_total)}`,`${m.standard_correct}/${m.standard_total} · ${pct(m.standard_correct/m.standard_total)}`,`${m.hard_correct}/${m.hard_total} · ${pct(m.hard_correct/m.hard_total)}`,String(m.rejected)]))} | |
| function renderDiagnostics(){let a=DIAG.screens.slice().sort((x,y)=>y.s800-x.s800);$('screenChart').innerHTML=a.map(x=>`<div class="diag-bar-row"><span title="${esc(x.change)}">${esc(x.name)}</span><div class="diag-track" title="800-update dev screen ${x.s800.toFixed(2)}%"><div class="diag-fill" style="width:${x.s800}%"></div></div><strong>${x.s800.toFixed(2)}%</strong></div>`).join('');makeTable($('screenTable'),['Experiment','Change tested','400 updates','800 updates','Full development'],DIAG.screens.map(x=>[`${esc(x.name)} <span class="dataset-sub">${x.id}</span>`,esc(x.change),x.s400.toFixed(2)+'%',x.s800.toFixed(2)+'%',x.fullDev==null?'Not run':x.fullDev.toFixed(2)+'%']));makeTable($('kevTable'),['Model','Suite','Partition','Rows','Accuracy'],DIAG.kevDev.map(x=>[esc(x.model),esc(x.suite),'Development',num(x.rows),x.score.toFixed(1)+'%']));makeTable($('secondaryTable'),['Model','Suite','Examples','Accuracy','Suite macro','Note'],DIAG.secondary.map(x=>[esc(x.model),esc(x.suite),num(x.rows),x.score.toFixed(2)+'%',pct(x.macro/100),esc(x.note)]))} | |
| document.querySelectorAll('[data-scope]').forEach(b=>b.addEventListener('click',()=>{state.scope=b.dataset.scope;document.querySelectorAll('[data-scope]').forEach(x=>x.setAttribute('aria-pressed',x===b));render()}));document.querySelectorAll('[data-metric]').forEach(b=>b.addEventListener('click',()=>{state.metric=b.dataset.metric;document.querySelectorAll('[data-metric]').forEach(x=>x.setAttribute('aria-pressed',x===b));render()})); | |
| renderDiagnostics();renderLocalJevbench();render();if(typeof window!=='undefined'){let resizeTimer;window.addEventListener('resize',()=>{clearTimeout(resizeTimer);resizeTimer=setTimeout(()=>renderDatasetDetail(sortModels(supportedModels())),100)})} | |
| </script> | |
| </body></html> | |