Add compact navigation, numbered descriptions, and the LightOn resume as example 3

.gitattributes

@@ -47,3 +47,4 @@ assets/cavailles-page-42.png filter=lfs diff=lfs merge=lfs -text
 assets/example2.png filter=lfs diff=lfs merge=lfs -text
 assets/transformer-page-9.png filter=lfs diff=lfs merge=lfs -text
 assets/tsla-q1-page-22.png filter=lfs diff=lfs merge=lfs -text
+assets/resume-lighton.png filter=lfs diff=lfs merge=lfs -text

assets/resume-lighton.json

@@ -0,0 +1,481 @@
+{
+  "page": 1,
+  "width": 1449,
+  "height": 2048,
+  "coordinates": "x1,y1,x2,y2 normalised to 0-1000",
+  "blocks": [
+    {
+      "label": "title",
+      "bbox": [
+        326,
+        38,
+        529,
+        82
+      ],
+      "text": "# LightOn",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        330,
+        87,
+        893,
+        105
+      ],
+      "text": "Open research lab | Document intelligence & retrieval, Paris",
+      "continues": false
+    },
+    {
+      "label": "text",
+      "bbox": [
+        331,
+        116,
+        931,
+        159
+      ],
+      "text": "Optical-computing startup turned open AI lab. We train the encoders, retrievers, rerankers and OCR models that make enterprise documents searchable, and release weights, papers, training data and ablation checkpoints under Apache 2.0.",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        24,
+        187,
+        104,
+        203
+      ],
+      "text": "## Contact",
+      "continues": false
+    },
+    {
+      "label": "list",
+      "bbox": [
+        48,
+        219,
+        262,
+        329
+      ],
+      "text": "- lighton.ai\n- Paris, France\n- hf.co/lightonai\n- github.com/lightonai\n- company/lighton\n- @LightOnIO\n- Euronext Growth Paris: ALTAI",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        23,
+        335,
+        102,
+        355
+      ],
+      "text": "## Origins",
+      "continues": false
+    },
+    {
+      "label": "image",
+      "bbox": [
+        26,
+        368,
+        123,
+        396
+      ],
+      "text": "logo \"ENS PSL UNIVERSITÉ PARIS\"",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        25,
+        403,
+        174,
+        418
+      ],
+      "text": "### Deep-tech spin-out",
+      "continues": false
+    },
+    {
+      "label": "text",
+      "bbox": [
+        23,
+        418,
+        261,
+        487
+      ],
+      "text": "Paris, physics & optics labs · 2016 Founded by I. Carron, L. Daudet, F. Krzakala and S. Gigan. Roots in optics and statistical physics at ENS Paris.",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        25,
+        492,
+        200,
+        507
+      ],
+      "text": "### Optical Processing Unit",
+      "continues": false
+    },
+    {
+      "label": "text",
+      "bbox": [
+        24,
+        507,
+        277,
+        573
+      ],
+      "text": "First product; on the Jean Zay supercomputer · 2018–21 Random projections in light: a non-von-Neumann co-processor for AI and HPC.",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        24,
+        577,
+        185,
+        590
+      ],
+      "text": "### IPO, Euronext Growth",
+      "continues": false
+    },
+    {
+      "label": "text",
+      "bbox": [
+        24,
+        590,
+        236,
+        618
+      ],
+      "text": "First listed GenAI company in Europe · Nov 2024",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        22,
+        624,
+        85,
+        644
+      ],
+      "text": "## Skills",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        23,
+        657,
+        99,
+        673
+      ],
+      "text": "### Research",
+      "continues": false
+    },
+    {
+      "label": "text",
+      "bbox": [
+        24,
+        674,
+        273,
+        742
+      ],
+      "text": "Modern encoders, late interaction, multilingual & long-context retrieval, LLM rerankers, end-to-end OCR & layout, RL with verifiable rewards, data curation, open releases",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        23,
+        746,
+        74,
+        761
+      ],
+      "text": "### Stack",
+      "continues": false
+    },
+    {
+      "label": "text",
+      "bbox": [
+        24,
+        762,
+        272,
+        801
+      ],
+      "text": "PyLate, pylate-rs, FastPlaid, NextPlaid, ColGREP, PyTorch, Sentence Transformers, vLLM, Rust",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        24,
+        805,
+        131,
+        821
+      ],
+      "text": "### Hugging Face",
+      "continues": false
+    },
+    {
+      "label": "text",
+      "bbox": [
+        24,
+        820,
+        273,
+        848
+      ],
+      "text": "89 models, 87 datasets, 17 collections; 2.8B released training pairs",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        24,
+        851,
+        112,
+        866
+      ],
+      "text": "### Languages",
+      "continues": false
+    },
+    {
+      "label": "text",
+      "bbox": [
+        23,
+        866,
+        276,
+        894
+      ],
+      "text": "French, English; models in 9 languages plus code",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        329,
+        196,
+        438,
+        214
+      ],
+      "text": "## Experience",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        888,
+        196,
+        955,
+        212
+      ],
+      "text": "### LightOn",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        330,
+        231,
+        797,
+        246
+      ],
+      "text": "#### Document intelligence & open retrieval · LightOn Research",
+      "continues": false
+    },
+    {
+      "label": "text",
+      "bbox": [
+        876,
+        233,
+        954,
+        244
+      ],
+      "text": "2026 – now",
+      "continues": false
+    },
+    {
+      "label": "list",
+      "bbox": [
+        334,
+        249,
+        935,
+        364
+      ],
+      "text": "- LightOnOCR-3 (0.8–4B): OCR, layout boxes, chart tables and captions, one model.\n- LightOnOCR-2: 1B end-to-end OCR, SOTA on OlmOCR-Bench, 250k downloads/month.\n- DenseOn, LateOn and multilingual twins: SOTA for size on BEIR, MIRACL, MLDR.\n- ColBERT-Zero: ColBERT pre-trained on public data only; beats GTE-ModernColBERT.\n- LightOn-rerank: multimodal rerankers (0.8–4B); the 4B beats an 8B on ViDoRe V3.\n- Agent-ModernColBERT: +10% on BrowseComp-Plus at 149M. NextPlaid, ColGREP.",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        330,
+        371,
+        769,
+        387
+      ],
+      "text": "#### Modern encoders & late interaction · LightOn Research",
+      "continues": false
+    },
+    {
+      "label": "text",
+      "bbox": [
+        870,
+        374,
+        954,
+        387
+      ],
+      "text": "2024 – 2025",
+      "continues": false
+    },
+    {
+      "label": "list",
+      "bbox": [
+        334,
+        389,
+        931,
+        460
+      ],
+      "text": "- ModernBERT with Answer.AI (ACL 2025): 8k context, 20M+ downloads.\n- PyLate (CIKM 2025) and FastPlaid: train and serve ColBERT models in a few lines.\n- GTE-ModernColBERT (SOTA, LongEmbed); Reason-ModernColBERT (SOTA ≤7B, BRIGHT).\n- MonoQwen2-VL, first visual document reranker. Ettin, BioClinical ModernBERT.",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        330,
+        466,
+        737,
+        483
+      ],
+      "text": "#### Enterprise LLMs, French and protein LMs · LightOn",
+      "continues": false
+    },
+    {
+      "label": "text",
+      "bbox": [
+        870,
+        469,
+        954,
+        482
+      ],
+      "text": "2019 – 2024",
+      "continues": false
+    },
+    {
+      "label": "list",
+      "bbox": [
+        334,
+        484,
+        947,
+        529
+      ],
+      "text": "- Alfred-40B (Falcon + RLHF), Mambaoutai (state-space LM), Paradigm platform, IPO.\n- PAGnol (LREC 2022), then largest French LM; RITA protein LMs up to 1.2B.\n- Direct Feedback Alignment at scale (NeurIPS 2020); OPU on Jean Zay.",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        330,
+        564,
+        468,
+        581
+      ],
+      "text": "## Open research",
+      "continues": false
+    },
+    {
+      "label": "table",
+      "bbox": [
+        330,
+        594,
+        952,
+        670
+      ],
+      "text": "<table>\n  <tr>\n    <td>LightOnOCR-3</td>\n    <td>0.8B, 1B, 4B · OCR + grounded layout in one model, RL-tuned</td>\n  </tr>\n  <tr>\n    <td>LightOn-rerank</td>\n    <td>Pointwise and listwise rerankers for text and page images</td>\n  </tr>\n  <tr>\n    <td>mDenseOn & mLateOn</td>\n    <td>307M multilingual retrievers, 8k context · 2.8B pairs released</td>\n  </tr>\n  <tr>\n    <td>ColBERT family</td>\n    <td>ColBERT-Zero, LateOn, LateOn-Code (17M–149M), Agent-ModernColBERT</td>\n  </tr>\n</table>",
+      "continues": false
+    },
+    {
+      "label": "text",
+      "bbox": [
+        331,
+        672,
+        912,
+        698
+      ],
+      "text": "On huggingface.co/lightonai with papers, data and ablation checkpoints. Tooling on GitHub: PyLate 900×, NextPlaid & ColGREP 550×, FastPlaid 280×, pylate-rs.",
+      "continues": false
+    },
+    {
+      "label": "title",
+      "bbox": [
+        330,
+        733,
+        529,
+        751
+      ],
+      "text": "## Selected publications",
+      "continues": false
+    },
+    {
+      "label": "list",
+      "bbox": [
+        334,
+        770,
+        871,
+        798
+      ],
+      "text": "- **DenseOn with the LateOn**: fully open dense and late-interaction models Sourty, Chaffin, Moura Junior, Chatelain. *Preprint 2026 arXiv*",
+      "continues": false
+    },
+    {
+      "label": "list",
+      "bbox": [
+        334,
+        798,
+        802,
+        826
+      ],
+      "text": "- **ColBERT-Zero**: to pre-train or not to pre-train ColBERT models Chaffin, Arnaboldi, Chatelain, Krzakala. *Preprint 2026 arXiv*",
+      "continues": false
+    },
+    {
+      "label": "list",
+      "bbox": [
+        334,
+        827,
+        871,
+        855
+      ],
+      "text": "- **LightOnOCR**: a 1B end-to-end multilingual VLM for state-of-the-art OCR Taghadouini, Cavaillès, Aubertin. *Preprint 2026 arXiv*",
+      "continues": false
+    },
+    {
+      "label": "list",
+      "bbox": [
+        334,
+        855,
+        812,
+        885
+      ],
+      "text": "- **PyLate**: flexible training and retrieval for late interaction models Chaffin, Sourty. *CIKM 2025 arXiv*",
+      "continues": false
+    },
+    {
+      "label": "list",
+      "bbox": [
+        334,
+        885,
+        910,
+        912
+      ],
+      "text": "- **Smarter, better, faster, longer**: a modern bidirectional encoder (ModernBERT) Warner, Chaffin, Clavié, Weller et al.. *ACL 2025 arXiv*",
+      "continues": false
+    },
+    {
+      "label": "text",
+      "bbox": [
+        331,
+        916,
+        933,
+        944
+      ],
+      "text": "Plus Layer Swap (*Findings of EMNLP 2026*), Ettin (2025), BioClinical ModernBERT (2025), RITA (2022), PAGnol (*LREC 2022*), DFA at scale (*NeurIPS 2020*). Field notes on hf.co/blog/lightonai.",
+      "continues": false
+    }
+  ]
+}

assets/resume-lighton.md

@@ -0,0 +1,127 @@
+![title](326,38,529,82) # LightOn
+
+![title](330,87,893,105) Open research lab | Document intelligence & retrieval, Paris
+
+![text](331,116,931,159) Optical-computing startup turned open AI lab. We train the encoders, retrievers, rerankers and OCR models that make enterprise documents searchable, and release weights, papers, training data and ablation checkpoints under Apache 2.0.
+
+![title](24,187,104,203) ## Contact
+
+![list](48,219,262,329)
+- lighton.ai
+- Paris, France
+- hf.co/lightonai
+- github.com/lightonai
+- company/lighton
+- @LightOnIO
+- Euronext Growth Paris: ALTAI
+
+![title](23,335,102,355) ## Origins
+
+![image](26,368,123,396) logo "ENS PSL UNIVERSITÉ PARIS"
+
+![title](25,403,174,418) ### Deep-tech spin-out
+
+![text](23,418,261,487) Paris, physics & optics labs · 2016 Founded by I. Carron, L. Daudet, F. Krzakala and S. Gigan. Roots in optics and statistical physics at ENS Paris.
+
+![title](25,492,200,507) ### Optical Processing Unit
+
+![text](24,507,277,573) First product; on the Jean Zay supercomputer · 2018–21 Random projections in light: a non-von-Neumann co-processor for AI and HPC.
+
+![title](24,577,185,590) ### IPO, Euronext Growth
+
+![text](24,590,236,618) First listed GenAI company in Europe · Nov 2024
+
+![title](22,624,85,644) ## Skills
+
+![title](23,657,99,673) ### Research
+
+![text](24,674,273,742) Modern encoders, late interaction, multilingual & long-context retrieval, LLM rerankers, end-to-end OCR & layout, RL with verifiable rewards, data curation, open releases
+
+![title](23,746,74,761) ### Stack
+
+![text](24,762,272,801) PyLate, pylate-rs, FastPlaid, NextPlaid, ColGREP, PyTorch, Sentence Transformers, vLLM, Rust
+
+![title](24,805,131,821) ### Hugging Face
+
+![text](24,820,273,848) 89 models, 87 datasets, 17 collections; 2.8B released training pairs
+
+![title](24,851,112,866) ### Languages
+
+![text](23,866,276,894) French, English; models in 9 languages plus code
+
+![title](329,196,438,214) ## Experience
+
+![title](888,196,955,212) ### LightOn
+
+![title](330,231,797,246) #### Document intelligence & open retrieval · LightOn Research
+
+![text](876,233,954,244) 2026 – now
+
+![list](334,249,935,364)
+- LightOnOCR-3 (0.8–4B): OCR, layout boxes, chart tables and captions, one model.
+- LightOnOCR-2: 1B end-to-end OCR, SOTA on OlmOCR-Bench, 250k downloads/month.
+- DenseOn, LateOn and multilingual twins: SOTA for size on BEIR, MIRACL, MLDR.
+- ColBERT-Zero: ColBERT pre-trained on public data only; beats GTE-ModernColBERT.
+- LightOn-rerank: multimodal rerankers (0.8–4B); the 4B beats an 8B on ViDoRe V3.
+- Agent-ModernColBERT: +10% on BrowseComp-Plus at 149M. NextPlaid, ColGREP.
+
+![title](330,371,769,387) #### Modern encoders & late interaction · LightOn Research
+
+![text](870,374,954,387) 2024 – 2025
+
+![list](334,389,931,460)
+- ModernBERT with Answer.AI (ACL 2025): 8k context, 20M+ downloads.
+- PyLate (CIKM 2025) and FastPlaid: train and serve ColBERT models in a few lines.
+- GTE-ModernColBERT (SOTA, LongEmbed); Reason-ModernColBERT (SOTA ≤7B, BRIGHT).
+- MonoQwen2-VL, first visual document reranker. Ettin, BioClinical ModernBERT.
+
+![title](330,466,737,483) #### Enterprise LLMs, French and protein LMs · LightOn
+
+![text](870,469,954,482) 2019 – 2024
+
+![list](334,484,947,529)
+- Alfred-40B (Falcon + RLHF), Mambaoutai (state-space LM), Paradigm platform, IPO.
+- PAGnol (LREC 2022), then largest French LM; RITA protein LMs up to 1.2B.
+- Direct Feedback Alignment at scale (NeurIPS 2020); OPU on Jean Zay.
+
+![title](330,564,468,581) ## Open research
+
+![table](330,594,952,670) <table>
+  <tr>
+    <td>LightOnOCR-3</td>
+    <td>0.8B, 1B, 4B · OCR + grounded layout in one model, RL-tuned</td>
+  </tr>
+  <tr>
+    <td>LightOn-rerank</td>
+    <td>Pointwise and listwise rerankers for text and page images</td>
+  </tr>
+  <tr>
+    <td>mDenseOn & mLateOn</td>
+    <td>307M multilingual retrievers, 8k context · 2.8B pairs released</td>
+  </tr>
+  <tr>
+    <td>ColBERT family</td>
+    <td>ColBERT-Zero, LateOn, LateOn-Code (17M–149M), Agent-ModernColBERT</td>
+  </tr>
+</table>
+
+![text](331,672,912,698) On huggingface.co/lightonai with papers, data and ablation checkpoints. Tooling on GitHub: PyLate 900×, NextPlaid & ColGREP 550×, FastPlaid 280×, pylate-rs.
+
+![title](330,733,529,751) ## Selected publications
+
+![list](334,770,871,798)
+- **DenseOn with the LateOn**: fully open dense and late-interaction models Sourty, Chaffin, Moura Junior, Chatelain. *Preprint 2026 arXiv*
+
+![list](334,798,802,826)
+- **ColBERT-Zero**: to pre-train or not to pre-train ColBERT models Chaffin, Arnaboldi, Chatelain, Krzakala. *Preprint 2026 arXiv*
+
+![list](334,827,871,855)
+- **LightOnOCR**: a 1B end-to-end multilingual VLM for state-of-the-art OCR Taghadouini, Cavaillès, Aubertin. *Preprint 2026 arXiv*
+
+![list](334,855,812,885)
+- **PyLate**: flexible training and retrieval for late interaction models Chaffin, Sourty. *CIKM 2025 arXiv*
+
+![list](334,885,910,912)
+- **Smarter, better, faster, longer**: a modern bidirectional encoder (ModernBERT) Warner, Chaffin, Clavié, Weller et al.. *ACL 2025 arXiv*
+
+![text](331,916,933,944) Plus Layer Swap (*Findings of EMNLP 2026*), Ettin (2025), BioClinical ModernBERT (2025), RITA (2022), PAGnol (*LREC 2022*), DFA at scale (*NeurIPS 2020*). Field notes on hf.co/blog/lightonai.

manifest.json

@@ -3,8 +3,9 @@
   "example_ids": [
     "transformer-page-9",
     "cavailles-page-42",
-    "example2",
+    "resume-lighton",
     "tsla-q1-page-22",
+    "example2",
     "td1-correction",
     "img-6981",
     "image-20",
@@ -32,6 +33,9 @@
     "assets/pxl-090958386.json": "38a9dbd7c75f0e2c215b6217413613786268c18393bb2c2331e7e03e25ec57d0",
     "assets/pxl-090958386.md": "dc24ee41673cb53e5d73acaff1f292a689c9353208e3e4971bf0502552289958",
     "assets/pxl-090958386.png": "517062933df17aa9a903af1eafdd634ce5edc0b179d0f52516e5a8cb0cc17e9d",
+    "assets/resume-lighton.json": "6e771439d21496f45043720b88638c9359f97a0055b898f52fd7f1b6f8e92ebc",
+    "assets/resume-lighton.md": "57bfd07467f0dd38070628d2d3e60103c1c0ad11b293f5b3af34b3d9e667a390",
+    "assets/resume-lighton.png": "082fe421272a9aca16af05a8f613d73d313cebd5c1d4230b3e27aa2744bda348",
     "assets/td1-correction.json": "14e7a2b9e2a4d76276459c3a157482653891ada44b2a27e74ba81298092f8fd6",
     "assets/td1-correction.md": "63b0ff92fa74f48dc9f88c09d9081e493a653ad26967c430d46b3edc071a998b",
     "assets/td1-correction.png": "95dcf292e1a503e22cdc425569ddfbe34b92056817c14bdc32a2a3baaf528023",
@@ -114,70 +118,77 @@
   "sources": [
     {
       "id": "transformer-page-9",
-      "title": "Transformer — page 9",
+      "title": "Example 1: Transformer table",
       "source_file": "transformer.pdf",
       "source_folder": "transformer",
       "source_page": 9
     },
     {
       "id": "cavailles-page-42",
-      "title": "Cavailles_Adrien_2019 — page 42",
+      "title": "Example 2: Quantum physics",
       "source_file": "CAVAILLES_Adrien_2019.pdf",
       "source_folder": "CAVAILLES_Adrien_2019-2",
       "source_page": 42
     },
     {
-      "id": "example2",
-      "title": "example2",
-      "source_file": "example2.pdf",
-      "source_folder": "example2",
+      "id": "resume-lighton",
+      "title": "Example 3: Company resume",
+      "source_file": "resume-lighton.pdf",
+      "source_folder": "resume-lighton",
       "source_page": 1
     },
     {
       "id": "tsla-q1-page-22",
-      "title": "TSLA-Q1 — page 22",
+      "title": "Example 4: Quarterly metrics",
       "source_file": "TSLA-Q1-2024-Update_1_1_ZagSVOv.pdf",
       "source_folder": "TSLA-Q1-2024-Update_1_1_ZagSVOv",
       "source_page": 22
     },
+    {
+      "id": "example2",
+      "title": "Example 5: Neutrino physics",
+      "source_file": "example2.pdf",
+      "source_folder": "example2",
+      "source_page": 1
+    },
     {
       "id": "td1-correction",
-      "title": "TD1-correction_exo1",
+      "title": "Example 6: Database exercise",
       "source_file": "TD1-correction_exo1.jpeg",
       "source_folder": "TD1-correction_exo1",
       "source_page": 1
     },
     {
       "id": "img-6981",
-      "title": "IMG_6981",
+      "title": "Example 7: Handwritten notes",
       "source_file": "IMG_6981.jpg",
       "source_folder": "IMG_6981",
       "source_page": 1
     },
     {
       "id": "image-20",
-      "title": "image-20",
+      "title": "Example 8: Calculus textbook",
       "source_file": "image (20).png",
       "source_folder": "image-20",
       "source_page": 1
     },
     {
       "id": "bastille",
-      "title": "Storming_the_bastille_4",
+      "title": "Example 9: Historical print",
       "source_file": "Storming_the_bastille_4.jpg",
       "source_folder": "Storming_the_bastille_4",
       "source_page": 1
     },
     {
       "id": "toolformer-page-2",
-      "title": "Toolformer — page 2",
+      "title": "Example 10: Research paper",
       "source_file": "Toolformer Language Models Can Teach Themselves to Use Tools.pdf",
       "source_folder": "Toolformer-Language-Models-Can-Teach-Themselves-to-Use-Tools",
       "source_page": 2
     },
     {
       "id": "pxl-090958386",
-      "title": "PXL_20260119_090958386",
+      "title": "Example 11: Geometry notes",
       "source_file": "PXL_20260119_090958386.jpg",
       "source_folder": "PXL_20260119_090958386",
       "source_page": 1