Add Output

• Architecture: Multi-Latent Attention, Peri-LN Transformer, Sliding Window Attention
• Training: GRPO, SnapPO, sDPO, RLVR, cascade distillation
• Efficiency: Layer-Adaptive Expert Pruning, Multi-Token Prediction
• Data: curriculum training, synthetic data pipelines

Prior innovations used — established techniques the model builds on:

• Attention: GQA, MLA, FlashAttention, sparse attention (NSA/DSA)
• Architecture: MoE, model merging, state space models (Mamba), GatedDeltaNets, Hyper-Connections
• Training: WSD/midtraining, muP scaling, DUS, FIM
• RL: GRPO, DPO, RLHF, RLVR, process reward models
• Efficiency: FP8 quantization, speculative decoding, knowledge distillation

Benchmarks — extract exact scores:

• General: MMLU, MMLU-Pro, Arena Hard
• Math: MATH-500, GSM8K, AIME 2024/2025
• Code: HumanEval, MBPP, SWE-Bench, LiveCodeBench
• Reasoning: GPQA-Diamond, ARC
• Domain-specific: KMMLU, tau-squared-Bench, etc.

For Models Without Papers (Blog/API-Only)

Many models (especially proprietary ones) are announced via blog posts without arxiv papers. In this case:

• Check the official blog post, announcement page, or API docs for specs
• Check HuggingFace model cards for architecture details
• Note "undisclosed" in description for unknown parameter counts
• Do NOT guess parameters — only include model.parameters if confirmed by a primary source
• Still check AA and OpenRouter for available data

External Links

For every model output, check:

• Artificial Analysis: Search site:artificialanalysis.ai [model name]. Fetch the page and extract the Intelligence Index score. Check for reasoning/adaptive variants — many models have multiple AA entries (e.g., model-name, model-name-reasoning, model-name-adaptive). Use the highest score and link to that variant's page.
• OpenRouter: Search site:openrouter.ai [model name]. Add the canonical model URL (without date suffix).
• HuggingFace model page: Find model weights (e.g., huggingface.co/org/model)
• HuggingFace blog: Check for technical blog posts at huggingface.co/blog/[org]/[post-slug]. These often contain detailed benchmarks, architecture explanations, and usage guides not found in the model card. Search site:huggingface.co/blog [model name].
• GitHub: Find code repo

For Closed Models with Technical Reports

Some frontier models (GPT-4, Claude, Gemini) are closed-source but have published technical reports:

• Use type: paper (not type: model) since there are no downloadable weights
• Still include AA Intelligence Index and OpenRouter links if the model is accessible via API
• Focus the description on what the technical report reveals: architecture choices, training scale, benchmark results, novel techniques
• Note what is NOT disclosed (e.g., "parameter count undisclosed", "training data composition not published")

For Foundational Technique Papers

Papers that introduced techniques now used industry-wide (e.g., "Attention Is All You Need", CLIP, InstructGPT/RLHF, Chain-of-Thought, DPO):

• Use type: paper
• The description should emphasize the technique's lasting impact, not just the original results
• Note which current models use the technique (e.g., "Transformers now underpin virtually all LLMs")
• Include citation count if exceptionally high (10K+) to signal importance

For Papers, Libraries, Datasets

• Fetch the arxiv page for title, authors, date, description
• SSRN papers: Some labs publish on SSRN (papers.ssrn.com) instead of arxiv, especially for economics/policy/interdisciplinary AI research. SSRN blocks automated fetching via Cloudflare — if you encounter this, ask the user to share the paper title/authors directly. Use the SSRN abstract URL as the source link, and paper: (with no arxiv ID) for the structured field.
• For libraries: find GitHub stars count, key features
• For datasets: find size, composition, intended use

For Scientific Foundation Models

We track scientific models that demonstrate frontier-level capabilities in scientific domains:

• Large-scale scientific foundation models with evidence of transfer learning across tasks (e.g., protein structure, materials discovery, weather prediction, genomics)
• Sci-LLMs that augment general LLMs with domain-specific tokenizers (e.g., SMILES for chemistry, amino acid sequences), additional modalities (molecular graphs, crystal structures, genomic sequences), specialized scientific knowledge, scientific reasoning chains, or agentic scientific workflows
• Note the scientific domain, any specialized tokenization/encoding, training data sources (PubMed, patents, experimental data), and whether the model demonstrates transfer to new tasks
• Examples: PanGu-Weather (Huawei), Matlantis/PFP (PFN), EXAONE Path (LG), BioNeMo (NVIDIA), AlphaFold (DeepMind)

2. Create Output YAML

Create data/outputs/{lab-slug}/{output-slug}.yaml: