` joined by ` ` (an en space) on a single source line. Each badge is 20px tall with 3px corners; the SVGs are 103, 97, 124 and 152px wide and the row spans about 515px. Details in Badges.
3. **Section nav** (y≈254): `` holding four `
` with seven `Installation •` style links (Installation, Quick Start, Dataset, Training, Evaluation, Pipeline, Citation), one per source line, separated by a bullet `•`. Body size, GitHub link blue.
4. **Rule** (y≈292): `---`, GitHub's 4px #d1d9e0 bar.
5. **First section** (y≈335): `## Overview`, ruled h2, then the summary paragraph with **SpanUQ** in bold and *span level* in italics.
6. **Results table** (from y≈672): `### Results` at y≈647 and a five-column table whose header and first four rows sit inside the first 900px.
There is no logo, subtitle, author line, affiliation line, teaser or caption.
## Badges
All four use the older `static/v1?label=&message=&color=` endpoint with no `style=`, so they render in shields.io's default `flat` style: 20px tall, 3px corner radius, a faint vertical gloss, Verdana 11px, label on #555555 and white (#ffffff) text with a #010101 text shadow, as the returned SVGs show.
| Label | Message | Color (hex) | Style | Links to |
|---|---|---|---|---|
| Project (github-pages logo) | Page | #007ec6 (`blue`; label #555555) | flat (default) | damon-demon.github.io/SpanUQ.html |
| ArXiv (arxiv logo) | Paper | #8b0000 (`darkred`, a CSS keyword the SVG writes as `fill="darkred"`; label #555555) | flat (default) | arxiv.org/abs/2607.05721 |
| 🤖 Released | Models | #67ac09 (`green`; label #555555) | flat (default) | huggingface.co/DamonDemon/SpanUQ |
| 🤗 Hugging Face | Dataset | #ea7233 (`orange`; label #555555) | flat (default) | huggingface.co/datasets/DamonDemon/SpanUQ-Benchmark |
| (nav row) | Installation • Quick Start • Dataset • Training • Evaluation • Pipeline • Citation | GitHub link blue | plain anchor links separated by `•` | in-page anchors |
The label names the thing and the message names the kind of destination, so the pattern is reversed on the two Hugging Face badges (label "🤖 Released", message "Models"). The named colors `green` and `orange` were read from the SVGs shields.io returns today (#67ac09 and #ea7233), which differ from the older shields.io values (#97ca00 and #fe7d37); the README itself only says `green` and `orange`.
No other colors are chosen by the author. The sample output block uses the emoji 🟢 🟡 🔴 as status dots, whose color comes from the viewer's emoji font. The rules are GitHub's #d1d9e0.
## Section Outline
| Level | Text | What it holds |
|---|---|---|
| h1 (centred, ruled) | [NeurIPS 2026] SpanUQ: Span-Level Uncertainty Quantification for LLM Generation | title; badge row, nav row and `---` follow |
| h2 | Overview | bold-led summary, "Key features:" and four bold-led bullets |
| h3 | Results | a five-column table, Model plus four metrics with ↑/↓ in the headers |
| h2 | Installation | a `bash` clone-and-install block, a requirements list, two one-line optional `bash` installs |
| h2 | Quick Start | h3 Step 1: Download Checkpoint & Dataset; h3 End-to-End Inference (Fastest Demo) with a sample-output block; h3 Step 2: Extract Hidden States (requires GPU); h3 Step 3: Evaluate with an "Expected output" block; h3 Alternative: On-the-fly Extraction; h3 Option B: Build Dataset from Scratch |
| h2 | Dataset | one sentence with a bold dataset name, a split-size table; h3 Data Format with two field tables; h3 Preparing Data for Training with a `bash` block and a directory tree |
| h2 | Training | a `bash` block, a Key training arguments table, a two-phase bullet list; h3 Post-hoc Calibration |
| h2 | Evaluation | a `bash` block, an Available methods table; h3 Layer Probing |
| h2 | Data Construction Pipeline | a Note blockquote, a four-step text diagram; h3 Prerequisites and h3 Step 1 to Step 4, each a `bash` block |
| h2 | Project Structure | an annotated tree in an untagged code block |
| h2 | Resources | three bullets with 📄 and 🤗 and bold labels |
| h2 | Citation | a `bibtex` block |
| h2 | License | one sentence |
Absent from the source: logo, teaser or any image, author and affiliation lines, news, demo GIF or video, ` ` with four flat `static/v1` shields joined by ` `: Project/Page `blue` (#007ec6) with the github-pages logo, ArXiv/Paper `darkred` (#8b0000) with the arxiv logo, 🤖 Released/Models `green` (#67ac09), 🤗 Hugging Face/Dataset `orange` (#ea7233); then a centred row of anchor links separated by `•`, and a `---` rule."
2. "Open with `## Overview`: a two-sentence summary with the name in bold, 'Key features:' and four bullets that start with a bold phrase, then `### Results` with a table `| Model | Metric ↑ | Metric ↓ |` aligned `|:------|:-------:|:-----:|`, three decimals, no bold."
3. "Write Quick Start as `### Step n: Verb Phrase` headings, each with a one-line lead and a `bash` block with one flag per backslash-continued line, followed by an untagged code block showing the expected output; put arguments, fields and methods in left-aligned reference tables with names in inline code."
## Similar Pages
- readme-seohongpark-ogbench: also a tool-first README with package badges and bold-led feature bullets, but with a logo and an environment grid instead of a results table.
- readme-mdyao-photoagent: also uses tables to make the method scannable and badges right under the title, but with `for-the-badge` buttons and heavy imagery.
- readme-eku127-satnav: a fuller academic header (subtitle, authors, affiliations, teaser) in front of a similar runnable Quick Start.
## Quick Start
````markdown
Installation •
Quick Start •
Dataset •
Training •
Evaluation •
Citation
`, acknowledgements, contact section, star-history chart.
Present but unusual: the venue is bracketed inside the h1; the nav row duplicates GitHub's own outline menu; the Quick Start interleaves numbered steps with unnumbered alternatives, and "Option B" has no "Option A"; the Pipeline note links to `#option-1-use-pre-built-dataset-recommended`, an anchor that no heading produces; the clone command uses `github.com/DamonDemon/SpanUQ.git` while the repo lives at `damon-demon/SpanUQ` (the DamonDemon path returned 404 on 2026-10-28); the Dataset sentence says "~293K annotated spans" while its table lists 344,288 spans for each Qwen backbone.
## Components
**Bracketed venue title.** `[Venue Year] Name: Full Title` as one centred ``, letting the heading rule close the title.
**Small flat badge row.** Four `static/v1` badges on one source line separated by ` `; label #555555, messages #007ec6, #8b0000, #67ac09, #ea7233; two carry Simple Icons logos (github-pages, arxiv), two carry emoji in the label.
**Bullet nav bar.** A centred paragraph of seven anchor links separated by `•`, followed by a `---` rule.
**Bold-led bullets.** "Key features:" followed by four bullets that each start with a bold phrase (**DETR-style span decoder**, **Mixture of Beta (MoB)**, **10–20× faster**, **Generalizes across LLM families**).
**Results table.** `| Model | AUROC ↑ | MAE ↓ | ρ_span ↑ | ρ_seq ↑ |` with `|:------|:-------:|:-----:|:--------:|:-------:|`: model names left-aligned, metrics centred, three decimals, no bold best values. GitHub renders it about 480px wide, about half the column.
**Numeric tables.** The split table right-aligns every count (`------:`) and uses thousands separators (17,468; 344,288).
**Reference tables.** Field, argument and method tables use `|:------|:-----------|` (left-aligned), with names in inline code and a dash "—" for empty defaults.
**Step headings with code.** `### Step n: Verb Phrase` followed by a one-line lead and a `bash` block with backslash-continued flags, one per line, indented four spaces.
**Sample output.** Untagged code blocks showing what the tool prints: a framed report with ═ and ─ rules and 🟢 🟡 🔴 status dots, and a plain "Expected output" metrics table.
**Inline notes.** Plain blockquotes with a bold label, `> **GPU requirement**: ...` and `> **Note**: ...`, not GitHub alert syntax.
**Annotated tree.** Untagged code blocks of box-drawing trees (`├──`, `└──`) with `# comment` annotations aligned in a column.
**Resources list.** Bullets of the form `- 🤗 **Model Checkpoints**: [org/name](url)`.
**Citation and license.** `@article{zhang2026spanuq, ...}` with `journal={arXiv preprint arXiv:2607.05721}`; one license sentence.
## Imagery
The README has no images, GIFs or videos. Its only visual elements are the four badges, GitHub tables, and code blocks; the "image" of the method is the sample output block, which shows annotated spans with 🟢 🟡 🔴 dots next to uncertainty scores, and the four-line text diagram of the data pipeline. Nothing depends on a light background, so the page reads the same on both GitHub themes.
## Adapting to Other Content
- **A library.** Keep the title without brackets, the flat badge row (Docs, PyPI, License) and the bullet nav bar; keep Installation and Quick Start as they are.
- **A dataset.** Move Dataset ahead of Installation, keep the split table with right-aligned counts and the two field tables, and add a sample-record code block.
- **A paper repo with a teaser.** Insert one `width="100%"` figure (856px) between the nav bar and the `---` rule; keep the Results table directly under Overview.
- **A method with several backbones.** Keep one row per backbone in the Results table and one config file per backbone in the tree.
- **A shorter README.** Cut the nav bar when there are fewer than five `##` sections; GitHub's outline menu covers it.
## Do's and Don'ts
| Do | Don't |
|---|---|
| Put a compact results table inside the first screen | Push numbers below install instructions |
| Show sample output after the first runnable command | Leave users to guess what a command prints |
| Give every argument, field and method its own small table | Describe flags in running prose |
| Right-align numeric columns in data tables | Center long integers |
| Number steps consistently and pair alternatives (Option A / Option B) | Leave "Option B" without an "Option A" |
| Make every in-page anchor match a real heading | Link to an anchor no heading produces |
| Use the repo's real path in the clone command | Clone from a different owner name |
## Agent Prompt Guide
1. "Write a README header: `
[Venue Year] Name: Full Title
`, then one `[Venue Year] Name: Full Title of the Paper
