add-lab

name	add-lab
description	Add or substantially update an AI research lab in Lab Index, including its profile, logo, people, news, outputs, provider links, and README entry. Use for new-lab onboarding, major lab-profile enrichment, corporate reorganizations, and lab slug or name changes.

Add a New Lab

1. Deep Research

Conduct thorough research before creating any files. When parallel agents are available and authorized, divide the work into 2-3 independent research passes:

• Agent 1: Corporate details, URLs, people profiles, news
• Agent 2: All models and papers (chronological), arxiv/HuggingFace/GitHub

Use this prompt template, adapting for the specific lab:

Tell me everything you can about [LAB NAME] research. What is their flagship model family, how many releases have they done, what is the best intelligence score they have achieved on Artificial Analysis, when were they founded, how much funding have they raised (or what is their current market cap if public), what other important papers, code, or datasets have they released, who are the 3-5 most significant people who lead or contribute to their AI research (with Google Scholar and OpenReview links), what are their GitHub and HuggingFace organizations, and what is their OpenRouter provider page? Do extensive research, double check all of your assertions, then produce a report with no preamble.

Required Information

Corporate & Financial:

• Official name and common short name (for name field)
• Founding year (company, not just the AI lab)
• Region/country
• Type: corporate, startup, nonprofit, academic
• For public companies: stock exchange, ticker, IPO year, current market cap
• For startups: latest valuation, funding total, key investors

URLs (verify each exists):

• Official website, Wikipedia page
• GitHub org(s) — some labs have multiple (e.g., Naver: naver, naver-ai, clovaai)
• HuggingFace org(s)
• Artificial Analysis provider page: search site:artificialanalysis.ai [lab name]
• OpenRouter provider page: search site:openrouter.ai [lab name]

Research Outputs (search all six):

• arxiv: site:arxiv.org [lab name] and search for specific model names
• HuggingFace: site:huggingface.co [lab org name]
• GitHub: check repos in their org(s)
• Lab research page: Many labs publish research pages that list papers not easily found via arxiv keyword search (e.g., machinelearning.apple.com/research, research.ibm.com, www.amazon.science). Check these for technique papers at top venues (ICLR, ICML, NeurIPS, CVPR, ICCV) — they are easy to miss but often more impactful than model releases.
• Lab's own publications index: Most labs maintain a blog-category or publications page listing their accepted papers (e.g., sbintuitions.co.jp/blog/archive/category/Publications). Always check this — keyword search on arxiv often misses papers with titles that don't mention the lab.
• SSRN: Some labs (especially those doing economics, policy, or interdisciplinary AI research) publish on SSRN (papers.ssrn.com) rather than arxiv. Search site:papers.ssrn.com [lab name]. Note that SSRN blocks automated fetching via Cloudflare — you may need to ask the user to share the paper title/authors.

Verify Affiliation at the PDF Level — Not Just Author Names

Keyword search will return papers written by people who were at the lab or who never were. Before claiming a paper belongs to a lab, open the arXiv HTML or PDF and read the affiliation block on page 1. Verify the author has an @lab-domain email or the lab is listed next to their name.

Common traps:

• Researchers move between labs. A scholar's recent paper may list their new employer, not the lab you're researching. Conversely, papers finished at a lab can appear months after the author has left — the work was done at the lab, so it still counts.
• Student collaborations. Papers with a university student as first author and lab researchers as co-authors often have the student's university as the primary affiliation and the lab as a secondary one. These still count as lab papers if the lab is listed. A purely student-authored paper with no lab co-authors does not.
• Common names. "Sho Tanaka" at Lab X is not necessarily the same "Sho Tanaka" at Lab Y. Check the Scholar/OpenReview profile before merging.

When a key researcher has left the lab, record it on the people entry as role: Former Researcher (dates; now at Lab Y) — this is more useful than omitting them, because their past work still defines the lab's output.

For each model found: total/active parameters, architecture (dense/MoE), context window, training tokens, benchmark scores, AA intelligence score, OpenRouter availability

People (3-5 key figures): name, role, Google Scholar URL, OpenReview URL, personal website, former affiliations. Verify it's the right person (common names are tricky).

Research Verification

• Cross-reference parameter counts between arxiv papers, HuggingFace model cards, and press
• Verify market cap/valuation against financial data sources
• Confirm people are currently at the lab (check dates)
• Distinguish published (arxiv) vs reported (press) parameter counts

2. Download Logo

Best source for AI companies: LobeHub icons. This works more reliably than Wikipedia for newer companies:

curl -sL "https://raw.githubusercontent.com/lobehub/lobe-icons/refs/heads/master/packages/static-png/dark/{name}-color.png" -o public/logos/{slug}.png
file public/logos/{slug}.png   # Must say "PNG image data", not "HTML document"

Common names to try: mistral, deepseek, upstage, zhipu. Check the LobeHub repo if unsure of the filename.

Fallback: Wikimedia Commons.

1. Find the logo on the lab's Wikipedia page, click through to the file page
2. Note the path (e.g., /wikipedia/commons/2/2d/SK_Telecom_Logo.svg)
3. Construct: https://upload.wikimedia.org/wikipedia/commons/thumb/2/2d/SK_Telecom_Logo.svg/500px-SK_Telecom_Logo.svg.png
4. Try /wikipedia/commons/ first, then /wikipedia/en/
5. Do NOT add a User-Agent header

Fallback: HuggingFace org avatar. Labs that have an HF org nearly always set their logo as the avatar, and the avatar is already 200x200. Fetch the avatar URL from the HF API:

AVATAR_URL=$(curl -s "https://huggingface.co/api/organizations/{hf-org-slug}/overview" | jq -r .avatarUrl)
curl -sL "$AVATAR_URL" -o /tmp/avatar.img
convert /tmp/avatar.img -resize 200x200 -gravity center -background white -extent 200x200 public/logos/{slug}.png

The HF avatar is often served as WebP regardless of file extension, so convert is required to normalize to PNG. Use this when LobeHub and Wikimedia don't have the lab (common for recently-founded subsidiaries and non-English labs).

Always verify and resize:

file public/logos/{slug}.png          # Must be "PNG image data"
convert public/logos/{slug}.png -resize 200x200 -gravity center -background white -extent 200x200 public/logos/{slug}.png

Last resort — ask the user for a logo URL, or create a placeholder:

convert -size 200x200 xc:'#brand-color' -gravity center -pointsize 28 -fill white -annotate 0 'Name' public/logos/{slug}.png

3. Create Lab YAML

Create data/labs/{slug}.yaml. Reference similar labs (e.g., mistral.yaml for European startups, deepseek.yaml for Chinese startups, lg-ai-research.yaml for Korean corporates).

Lab Naming

Use the highest-level corporate owner as the lab name, not the AI division or model family:

• "Kakao" not "Kakao Brain" or "Kanana"
• "NVIDIA" not "NVIDIA Research" or "Nemotron"
• "Naver" not "NAVER AI Lab" or "HyperCLOVA"

For well-known abbreviations, use the short form: "PFN" not "Preferred Networks", "BAAI" not "Beijing Academy of Artificial Intelligence".

If the lab has been reorganized (e.g., Kakao Brain merged into Kakao Corp → Kanana formed), explain this history in the description — it helps readers understand the continuity of research.

Multiple GitHub/HuggingFace Orgs

Many labs have legacy orgs from before reorganization. The lab YAML only has one github: and one huggingface: field — use the current/primary org. Mention legacy orgs in the description or link them from individual output source lists:

• Kakao: primary kakaocorp on HF, legacy kakaobrain on HF/GitHub
• PFN: primary pfnet on HF, also pfnet-research on GitHub

name: Lab Name
slug: lab-slug
url: https://...
wikipedia: https://en.wikipedia.org/wiki/...
huggingface: https://huggingface.co/...
github: https://github.com/...
artificialanalysis: https://artificialanalysis.ai/providers/...
openrouter: https://openrouter.ai/provider-slug
region: country                  # china, korea, france, usa, etc.
founded: "YYYY"
type: startup                    # corporate, startup, nonprofit, academic
valuation:
  amount: "$14B"
  type: private                  # market-cap, private, or revenue
  date: "2025-09"
description: >
  <p>Paragraph 1: identity, founding, backing, scale (funding, revenue, team size).</p>
  <p>Paragraph 2: flagship model evolution (chronological, with links and numbers).</p>
  <p>Paragraph 3: other notable research, products, and what makes this lab distinctive.</p>
people:
  - name: Person Name
    url: https://personal-site.com/
    role: Current Role
    formerly: Previous Company   
    
         
         
        
         

    
     
     
     

   

Gathering News Items

Every lab should have 2-5 news items covering major milestones (launches, funding, partnerships). Search these sources in priority order:

1. Lab's own blog/newsroom — check first, most authoritative
2. Bloomberg / TechCrunch / CNBC / The Information — Tier 1 general sources
3. Regional sources by lab location:
   • China: SCMP, ChinaTalk, Caixin, LatePost (晚点), 36Kr, Pandaily
   • Japan: Nikkei Asia, lab press releases
   • Korea: Korea Times, Nikkei Asia
   • Israel: Calcalist
   • US/Europe: VentureBeat, GeekWire, Wired, Ars Technica, The Verge
4. AI-specialist sources: Latent Space, Interconnects — for technical context

See AGENTS.md "Priority News Sources" for the full tiered list.

New Region Check

If this is the first lab from a new country, add the country's flag emoji to the regionFlags map in src/pages/index.astro:

const regionFlags: Record<string, string> = {
  china: '\u{1F1E8}\u{1F1F3}',
  korea: '\u{1F1F0}\u{1F1F7}',
  france: '\u{1F1EB}\u{1F1F7}',
  // add new country here
};

Description Guidelines

• Write in HTML using <p>, <strong>, <a>, — within YAML > blocks
• Link to arxiv papers, GitHub repos, and other lab pages (<a href="/labs/deepseek">)
• Include concrete numbers: parameter counts, training tokens, benchmark scores, funding
• Tell a story — what makes this lab distinctive?
• Complex corporate structures: If a lab's region might surprise readers (e.g., Singapore entity with Chinese parent, or a multinational with labs in multiple countries), proactively explain the legal structure and why the lab is classified in its region. Include incorporation details, where the research team is based, and local institutional ties.

4. Create Outputs

mkdir -p data/outputs/{slug}

See /add-output for detailed format. Key decisions:

User-Directed Filtering

If the user specifies a focus (e.g., "focus on Nemotron, only include others if truly significant"), respect that. Prolific labs like NVIDIA or Google may have 50+ outputs but only 10-15 that matter for our tracking focus.

Our core interests (see AGENTS.md "Research Focus"):

• Frontier foundation models great at language, reasoning, coding, multi-turn agentic work, and multimodality
• Scientific foundation models of large scale with evidence of transfer learning, and Sci-LLMs augmenting LLMs with scientific tokenizers, modalities, knowledge, reasoning, and agentic uses
• Foundational technique papers that introduced methods now standard across labs
• Training infrastructure (frameworks, datasets, evaluation, data curation)

Ask: "Does this output advance the frontier of intelligence, enable scientific discovery, or introduce a technique that changed how models are built?"

Don't overlook scaling science and technique papers. Labs like Apple, Google, and Microsoft produce highly influential papers on scaling laws, optimizers, tokenization, and training methodology that are published at top venues (ICLR, ICML, NeurIPS, CVPR) but are easy to miss because they aren't "model releases." These are often more impactful than any single model — e.g., AdEMAMix (Apple, ICLR 2025) is an optimizer adopted across the field, and Scaling Laws for Native Multimodal Models (Apple, ICCV 2025 Oral) studied 457 models. Search the lab's research page and Google Scholar profiles of key researchers, not just arxiv keyword searches for model names.

Exclude from the start:

• Business-specific applications (delivery logistics, recommendation systems, e-commerce search, customer service, mobile infra)
• Narrow benchmarks/datasets that didn't become industry standards
• Minor papers with low community adoption (<500 GitHub stars, no notable citations) unless science-related or first-in-series

Closed-Source / Frontier Labs (Google, OpenAI, Anthropic, Meta, xAI)

Some of the most important labs primarily release closed models. For these labs:

Models to include:

• Models with published technical reports (GPT-4, Claude 3, Gemini, PaLM) — even without open weights, the paper is the output
• Open-weight releases from otherwise closed labs (Gemma from Google, Llama from Meta) — these are especially important
• Models with Artificial Analysis scores — even closed models can be tracked by their AA Intelligence Index

Foundational research papers — the most important outputs. These labs often produced the techniques that ALL other labs now use. Search deeply for papers that introduced or popularized:

Architecture:

• Attention mechanisms: self-attention, multi-head attention, Transformer ("Attention Is All You Need"), grouped-query attention (GQA), multi-query attention (MQA), multi-head latent attention (MLA), sliding window attention, Flash Attention, ring attention, linear attention
• Mixture of Experts: Switch Transformer, GShard, fine-grained MoE
• State space models: S4, Mamba, Mamba-2
• Hybrid architectures: Transformer+SSM combinations
• Position encodings: RoPE, ALiBi, NTK-aware scaling, YaRN

Tokenization & pre-training:

• BPE, SentencePiece, Unigram tokenization
• Pre-training objectives: causal LM, masked LM, span corruption, fill-in-the-middle (FIM), blank infilling
• Scaling laws: Kaplan et al., Chinchilla

Training methods:

• Optimizers: Adam, AdamW, Lion, Muon, AdEMAMix
• Learning rate schedules: cosine decay, WSD (warmup-stable-decay), cooldown
• Mixed precision: FP16, BF16, FP8, NVFP4
• Distributed training: tensor/pipeline/data parallelism, FSDP, ZeRO, Megatron-LM
• Curriculum learning, staged training, mid-training, depth up-scaling

Post-training & alignment:

• RLHF (InstructGPT), PPO for LLMs, Constitutional AI
• DPO, GRPO, RLVR, sDPO, process reward models
• Safety training, red-teaming methodologies

Reasoning & agents:

• Chain-of-Thought prompting, tree of thought
• O1-style reasoning / thinking tokens / test-time compute scaling
• Tool use, function calling, ReAct, code execution as reasoning

Multimodality:

• CLIP (image-text alignment), ViT (Vision Transformer)
• Visual instruction tuning (LLaVA)
• Speech: Whisper, speech-text interleaving
• Diffusion models for generation

Interpretability & evaluation:

• Mechanistic interpretability, circuits, sparse autoencoders
• Benchmarks that became industry standards (MMLU, HumanEval, SWE-Bench, Chatbot Arena)

When adding these labs, prioritize papers by lasting impact — a technique paper cited 10,000+ times that changed how everyone trains models is more important than a closed model that was SOTA for 3 months.

What Gets Its Own Output Page

Search the lab's FULL history, not just current frontier models. Outputs from earlier eras often define the lab's identity and influence (e.g., Google's Transformer paper, OpenAI's CLIP, Kakao's KoGPT, PFN's Chainer). Check the lab's GitHub and HuggingFace for older repos with high star counts.

Create an output for:

• Each major model family or version (Nemotron-4 340B, Nemotron-H, Nemotron 3 Super)
• Each distinct product line (Codestral, Pixtral, Cosmos)
• The lab's first significant model (establishes their entry into the field)
• Foundational technique papers that introduced methods now used industry-wide
• Widely-used open-source tools/libraries (Megatron-LM, Optuna, Chainer, vLLM)
• Significant datasets (COYO-700M, Dolma, RefinedWeb, Darwin-CC, The Pile)
• Evaluations/benchmarks (type: eval) that became industry standards or are used in the AA Intelligence Index (10 evals) or Epoch Capabilities Index (~42 benchmarks). Examples: HumanEval (OpenAI), GPQA (Anthropic+NYU), RULER (NVIDIA), IFBench (AI2), BBEH (Google DeepMind), Belebele (Meta), GDPval (OpenAI). Check both composite indices to see if any of the lab's benchmarks are included.
• Training infrastructure and data research — data curation, scaling laws, data mixing, post-training methods, evaluation frameworks, reward modeling

Do NOT create separate outputs for:

• Point releases within the same version (v0.1, v0.2, v0.3 of the same model)
• Size variants of the same version (use model.variants instead)
• Instruct/Chat fine-tunes of a base model (note in the base model's description)
• Deprecated models superseded by a direct successor
• Announced but unreleased models (no output until weights or API are available)
• Closed models with no technical report or paper (nothing to document)

Flagship Criteria

flagship: true is a permanent landmark marker, judged at the time of release — a lab accumulates many over time (major labs have 10–30+), covering every headline model generation plus seminal papers/artifacts. It does not mean "the single current best model."

Mark flagship: true when the output was, at release, a step change for the lab:

• New architecture or scale milestone (first MoE, first 100B+, etc.)
• New capability (first multimodal, first reasoning, first code model)
• A headline model generation, or the best model in a product line at its time

When adding a new flagship, do NOT demote its predecessors — a model that was flagship-worthy at release stays flagged after a successor supersedes it (every model generation stays flagged side by side).

Do NOT mark as flagship: minor updates, size variants, specialized fine-tunes.

Proprietary Models with Undisclosed Details

Some models are API-only with no published parameter counts:

• Note "undisclosed" in the description if parameters are unknown
• Do not guess — only include structured model.parameters if confirmed
• Still include model.intelligence_index if AA has scored the model
• Still include OpenRouter links

From Scratch vs Derivative — Critical for Scale Column

The home page "Scale" column shows the largest model each lab trained from scratch. Only set model.parameters on models the lab pretrained themselves. For models fine-tuned/adapted from another lab's base (e.g., Qwen, Llama, Mistral), use model.base_model instead and omit parameters. See /add-output for detailed guidance and examples.

5. Add OpenRouter Links

Fetch the provider page (e.g., openrouter.ai/mistralai) and map models to outputs.

• Add openrouter: to the lab YAML
• For each output with a corresponding OpenRouter model, add an OpenRouter source link
• OpenRouter often has many more models than we have outputs (e.g., 42 vs 16 for Mistral). This is expected — we don't need an output for every OpenRouter variant.

6. Update README.md

• Increment the lab count
• Add the lab name to the appropriate country (alphabetical)
• If first lab from a new country, add a new country section
• Update the description if it says "Asian" and the lab is not from Asia

7. Validate and Build

npm run validate     # Check YAML schema and project invariants
npm run build        # Verify page generation and integration

YAML changes are picked up on browser refresh without restarting the dev server. Only restart if you changed .astro templates or .ts code.

8. Checklist

• Deep research completed with parallel agents
• Logo verified as PNG and resized to 200x200
• Lab YAML with all fields, 2-3 paragraph HTML description
• People with Scholar/OpenReview links (3-7 people)
• News items added (funding rounds, major launches, partnerships — don't skip this)
• Outputs created for major models/papers (not every minor variant)
• flagship: true only for genuine milestones
• Structured model fields (architecture, parameters, context_window) where confirmed
• AA intelligence scores where available
• OpenRouter links at lab and model level
• Region flag exists in index.astro for this country
• README.md updated
• npm run validate and npm run build pass

Updating an Existing Lab

1. Read the current file first — don't overwrite existing content
2. Enrich, don't replace — weave new info into existing paragraphs
3. Verify claims — search the web to confirm new facts
4. Valuation updates — update both amount and date
5. Slug/name renames — high-impact: requires renaming YAML, output dir, all lab: refs, logo, metrics.json, and cross-references