{"passport":{"unfragile":{"@version":"1.0","version":"2026-05","artifact":{"id":"awesome-gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox","slug":"gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox","name":"GPT-NeoX-20B: An Open-Source Autoregressive Language Model (GPT-NeoX)","type":"model","url":"https://arxiv.org/abs/2204.06745","page_url":"https://unfragile.ai/gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox","categories":["productivity"],"tags":[],"pricing":{"model":"unknown","free":false,"starting_price":null},"status":"inactive","verified":false},"capabilities":[{"id":"awesome-gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox__cap_0","uri":"capability://text.generation.language.autoregressive.text.generation.with.20b.parameters","name":"autoregressive text generation with 20b parameters","description":"Generates coherent multi-token sequences using a transformer-based autoregressive architecture with 20 billion parameters trained on 825GB of curated text data. Uses standard causal language modeling with next-token prediction loss, enabling generation of arbitrary-length outputs through iterative sampling or beam search. Implements efficient inference through batch processing and supports both greedy decoding and nucleus/top-k sampling strategies for controlling output diversity.","intents":["Generate long-form text content (articles, stories, code) from natural language prompts","Build conversational AI systems without proprietary API dependencies","Fine-tune a large pretrained model for domain-specific text generation tasks","Run inference on-premises or in air-gapped environments for sensitive applications"],"best_for":["Open-source ML practitioners building self-hosted language model applications","Organizations requiring full model transparency and control over training data","Researchers studying large language model behavior and interpretability","Teams with GPU infrastructure (A100, H100) seeking cost-effective alternatives to closed APIs"],"limitations":["20B parameters requires 40GB+ VRAM for full precision inference (16-bit requires 40GB, 8-bit quantization reduces to ~20GB but adds latency)","Inference speed significantly slower than optimized commercial APIs — ~50-100ms per token on A100 vs 10-20ms for GPT-3.5","Knowledge cutoff at April 2022; cannot access real-time information without external retrieval","No instruction-tuning or RLHF applied in base model — requires additional fine-tuning for chat/instruction-following tasks","Context window limited to 2048 tokens, insufficient for long-document analysis without chunking"],"requires":["GPU with minimum 40GB VRAM (A100 80GB, H100, or equivalent) for full precision","PyTorch 1.11+ or compatible deep learning framework","Python 3.8+","825GB disk space for full model weights (or 5-10GB for quantized versions)","CUDA 11.3+ for GPU acceleration (or CPU inference with 10-50x latency penalty)"],"input_types":["text prompts (natural language or structured templates)","partial sequences (for continuation/completion tasks)","token IDs (for low-level control)"],"output_types":["text sequences (variable length)","token probability distributions (for uncertainty quantification)","embeddings (hidden layer activations for downstream tasks)"],"categories":["text-generation-language","open-source-llm"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox__cap_1","uri":"capability://text.generation.language.instruction.following.and.chat.adaptation.through.fine.tuning","name":"instruction-following and chat adaptation through fine-tuning","description":"Provides a base model architecture optimized for downstream fine-tuning on instruction-following and conversational datasets. The model uses standard transformer blocks with rotary positional embeddings (RoPE) and parallel attention/MLP computation, enabling efficient adaptation to chat, Q&A, and task-specific behaviors through supervised fine-tuning (SFT) on curated instruction datasets. Supports parameter-efficient fine-tuning methods like LoRA for adapting the 20B model with <1GB additional parameters.","intents":["Fine-tune the base model on proprietary instruction datasets to create domain-specific assistants","Adapt the model for chat interfaces with minimal labeled data (1K-10K examples)","Create specialized versions for code generation, summarization, or question-answering","Implement parameter-efficient fine-tuning to reduce training compute and storage overhead"],"best_for":["ML teams with labeled instruction datasets and GPU clusters for fine-tuning","Organizations building vertical-specific AI assistants (legal, medical, financial)","Researchers studying instruction-tuning and alignment techniques","Companies wanting to customize model behavior without retraining from scratch"],"limitations":["Base model lacks instruction-tuning — raw outputs often verbose, unfocused, or off-topic without fine-tuning","Fine-tuning requires 10-100GB GPU memory depending on batch size and sequence length","No built-in safety alignment (RLHF) — fine-tuned models inherit base model's lack of refusal behaviors","Instruction-following quality depends heavily on fine-tuning dataset quality and size; small datasets (<1K examples) often underperform","LoRA adaptation adds inference latency (~5-10%) due to rank decomposition overhead"],"requires":["Labeled instruction dataset (minimum 500-1000 examples for meaningful adaptation)","GPU with 40GB+ VRAM for full fine-tuning, or 24GB+ for LoRA-based adaptation","PyTorch 1.11+ with distributed training support (torch.nn.parallel.DistributedDataParallel)","Fine-tuning framework (HuggingFace Transformers, DeepSpeed, or Megatron-LM)","Python 3.8+"],"input_types":["instruction-response pairs (JSON/CSV format)","conversation histories (multi-turn dialogue)","task-specific examples (code, summaries, Q&A)"],"output_types":["fine-tuned model weights (full or LoRA adapters)","evaluation metrics (BLEU, ROUGE, task-specific accuracy)","inference-ready model checkpoints"],"categories":["text-generation-language","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox__cap_2","uri":"capability://automation.workflow.multi.gpu.distributed.inference.with.model.parallelism","name":"multi-gpu distributed inference with model parallelism","description":"Supports efficient inference across multiple GPUs using tensor parallelism and pipeline parallelism strategies, enabling deployment of the 20B model on clusters of consumer/enterprise GPUs. Implements layer-wise partitioning where different transformer layers run on different devices, with optimized communication patterns to minimize inter-GPU bandwidth overhead. Integrates with DeepSpeed and Megatron-LM for production-grade distributed inference with dynamic batching.","intents":["Deploy the 20B model across multiple GPUs to achieve acceptable inference latency for production services","Serve multiple concurrent requests by batching inference across a GPU cluster","Reduce per-GPU memory requirements by distributing model weights across devices","Scale inference throughput horizontally without retraining or quantization"],"best_for":["Teams deploying language models as production APIs with SLA requirements","Organizations with multi-GPU infrastructure (2-8 GPUs) seeking to maximize throughput","Service providers offering hosted inference for the 20B model","Research labs studying distributed inference optimization"],"limitations":["Inter-GPU communication overhead (PCIe, NVLink, or network) adds 10-30ms latency per inference step depending on topology","Requires low-latency interconnect (NVLink preferred; PCIe 4.0+ acceptable; network-based parallelism impractical)","Setup complexity — requires careful tuning of partition boundaries and batch sizes for optimal performance","Not all fine-tuned adapters (LoRA) are compatible with distributed inference without custom integration","Scaling beyond 8 GPUs shows diminishing returns due to communication overhead dominating compute time"],"requires":["2-8 GPUs with 40GB+ VRAM each (A100, H100, or equivalent)","NVLink or PCIe 4.0+ interconnect for low-latency communication","DeepSpeed 0.7+ or Megatron-LM for distributed inference orchestration","NCCL 2.10+ for optimized GPU communication","Python 3.8+, PyTorch 1.11+"],"input_types":["text prompts (batched or streaming)","token IDs with attention masks","generation parameters (temperature, top-k, max_length)"],"output_types":["generated text sequences (batched)","token logits (for ensemble methods)","latency/throughput metrics"],"categories":["automation-workflow","tool-use-integration"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox__cap_3","uri":"capability://automation.workflow.quantization.aware.inference.8.bit.and.4.bit","name":"quantization-aware inference (8-bit and 4-bit)","description":"Enables reduced-precision inference through post-training quantization to 8-bit or 4-bit integer representations, reducing model size from 40GB to 10-20GB while maintaining 95%+ output quality. Uses symmetric quantization with learned scale factors per layer, implemented via libraries like bitsandbytes and GPTQ. Quantized models run on consumer GPUs (24GB VRAM) with 20-40% latency overhead compared to full precision, enabling broader deployment.","intents":["Deploy the 20B model on consumer GPUs (RTX 4090, RTX 3090) with limited VRAM","Reduce model storage and download time from 40GB to 10-20GB for edge deployment","Run inference on CPU or mobile devices with acceptable latency (via 4-bit quantization)","Reduce inference cost per token by 30-50% through smaller model footprint and faster memory access"],"best_for":["Individual developers and small teams without enterprise GPU access","Edge deployment scenarios (laptops, on-premises servers with limited VRAM)","Cost-sensitive production deployments where latency tolerance is 100-200ms","Research on quantization-aware training and low-bit model compression"],"limitations":["8-bit quantization adds 20-30% latency overhead; 4-bit adds 40-60% overhead due to dequantization operations","Output quality degrades measurably at 4-bit precision — hallucination rate increases 10-20%, factual accuracy drops 5-15%","Quantization is post-training — cannot improve model quality, only reduce compute cost","Not all fine-tuning methods are compatible with quantized models; LoRA adapters require special handling","Quantized models lose some numerical precision for long sequences (>1000 tokens), affecting coherence"],"requires":["bitsandbytes 0.35+ or GPTQ library for quantization","GPU with 24GB+ VRAM for 8-bit (RTX 4090, RTX 3090, A6000), or 12GB+ for 4-bit","PyTorch 1.11+","CUDA 11.3+ (or CPU-only mode with 5-10x latency penalty)","Python 3.8+"],"input_types":["full-precision model weights (40GB)","quantization configuration (bit-width, scale factors)","text prompts"],"output_types":["quantized model weights (10-20GB)","generated text (same format as full-precision)","quantization statistics (scale factors, clipping ranges)"],"categories":["automation-workflow","data-processing-analysis"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox__cap_4","uri":"capability://memory.knowledge.embedding.extraction.and.semantic.representation","name":"embedding extraction and semantic representation","description":"Extracts dense vector representations (embeddings) from intermediate transformer layers, enabling semantic search, clustering, and similarity-based retrieval tasks. Outputs embeddings from configurable layers (typically final hidden state or pooled representation) with 4096-dimensional vectors. Embeddings capture semantic meaning of input text and can be indexed in vector databases (Pinecone, Weaviate, Milvus) for efficient similarity search at scale.","intents":["Build semantic search systems over document collections using embedding similarity","Cluster documents or queries by semantic meaning for content organization","Implement retrieval-augmented generation (RAG) by finding relevant context via embedding similarity","Create recommendation systems based on semantic similarity between items"],"best_for":["Teams building semantic search or RAG systems with open-source models","Organizations needing embeddings for clustering or similarity tasks without API dependencies","Researchers studying embedding quality and semantic representation learning","Companies implementing recommendation systems with semantic matching"],"limitations":["Embeddings are 4096-dimensional — larger than typical embedding models (384-768 dims), increasing storage and search latency by 5-10x","Embeddings are not fine-tuned for specific domains — generic semantic representations may not capture domain-specific similarity","Embedding quality depends on pretraining data; knowledge cutoff at April 2022 limits relevance for recent concepts","No built-in embedding normalization or dimensionality reduction — requires external processing for optimal vector database performance","Batch embedding extraction requires GPU memory proportional to batch size; CPU embedding extraction is 50-100x slower"],"requires":["GPU with 40GB+ VRAM for batch embedding extraction, or 24GB+ for single-sample inference","Vector database (Pinecone, Weaviate, Milvus, Qdrant) for efficient similarity search","PyTorch 1.11+","Python 3.8+","Optional: dimensionality reduction library (UMAP, PCA) for visualization or optimization"],"input_types":["text sequences (variable length, up to 2048 tokens)","batches of documents or queries","layer indices (for selecting which transformer layer to extract from)"],"output_types":["dense vectors (4096-dimensional float32)","similarity scores (cosine, L2 distance)","nearest neighbor indices (for retrieval)"],"categories":["memory-knowledge","search-retrieval"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox__cap_5","uri":"capability://text.generation.language.few.shot.and.zero.shot.task.adaptation","name":"few-shot and zero-shot task adaptation","description":"Performs task adaptation through in-context learning by conditioning the model on a few examples (few-shot) or task descriptions (zero-shot) without parameter updates. The model uses its pretrained knowledge to infer task structure from examples and generate appropriate outputs. Supports various prompt formats (instruction-based, example-based, chain-of-thought) to guide model behavior for tasks not explicitly seen during training.","intents":["Adapt the model to new tasks (classification, extraction, summarization) with 1-10 examples without fine-tuning","Perform zero-shot task inference using natural language task descriptions","Implement chain-of-thought prompting to improve reasoning on complex tasks","Rapidly prototype task-specific applications without labeled training data"],"best_for":["Rapid prototyping teams building task-specific applications with minimal labeled data","Researchers studying in-context learning and prompt engineering","Organizations with diverse, changing tasks that don't justify fine-tuning","Developers building flexible NLP pipelines that adapt to new tasks dynamically"],"limitations":["Few-shot performance highly sensitive to example selection and prompt formatting — 10-30% variance depending on prompt engineering","Context window limited to 2048 tokens — cannot include many examples or long documents for in-context learning","Zero-shot performance significantly lower than fine-tuned models — 20-40% accuracy gap on structured tasks (classification, extraction)","No mechanism to learn task-specific patterns — each inference requires full context, increasing latency and cost","Prompt engineering requires domain expertise and iterative refinement; no automated prompt optimization"],"requires":["GPU with 40GB+ VRAM for inference (or 24GB+ with quantization)","Prompt engineering framework (optional: LangChain, Prompt Engineering Guide)","PyTorch 1.11+","Python 3.8+","Examples or task descriptions in natural language"],"input_types":["natural language task descriptions","few-shot examples (1-10 input-output pairs)","chain-of-thought prompts","structured prompt templates"],"output_types":["task-specific predictions (classification labels, extracted entities, summaries)","confidence scores (via logit analysis)","reasoning traces (for chain-of-thought)"],"categories":["text-generation-language","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox__cap_6","uri":"capability://code.generation.editing.code.generation.and.completion","name":"code generation and completion","description":"Generates and completes code across multiple programming languages (Python, JavaScript, C++, Java, etc.) using transformer-based autoregressive prediction trained on code-heavy portions of The Pile dataset. Supports both function-level completion (single function body) and file-level generation (multi-function modules). Implements standard code generation patterns including docstring-to-code, comment-to-code, and partial-code-to-completion.","intents":["Generate boilerplate code and function implementations from natural language descriptions","Complete partial code snippets with context-aware suggestions","Translate between programming languages using semantic understanding","Generate test cases and documentation from code"],"best_for":["Developers using open-source code generation without proprietary API dependencies","Teams building code-centric applications (IDE plugins, code review tools) with full model control","Organizations with code in restricted/proprietary domains requiring on-premises inference","Researchers studying code generation and program synthesis"],"limitations":["Code generation quality lower than specialized models (Codex, CodeLLaMA) — 30-50% lower pass rate on HumanEval benchmark","Limited to 2048-token context — cannot generate large files or maintain consistency across multiple functions","No syntax validation — generates syntactically invalid code 10-20% of the time; requires external linting","No type checking or semantic analysis — generated code may have runtime errors despite syntactic correctness","Knowledge cutoff at April 2022 — unfamiliar with recent language features and library APIs"],"requires":["GPU with 40GB+ VRAM for inference (or 24GB+ with quantization)","PyTorch 1.11+","Python 3.8+","Optional: linting/type-checking tools (pylint, mypy, ESLint) for validation","Optional: IDE plugin framework (VS Code Extension API, JetBrains Plugin SDK)"],"input_types":["natural language descriptions (docstrings, comments)","partial code snippets","function signatures","test cases or examples"],"output_types":["generated code (function bodies, modules, scripts)","completion suggestions (ranked by likelihood)","confidence scores (via logit analysis)"],"categories":["code-generation-editing","text-generation-language"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox__cap_7","uri":"capability://text.generation.language.multilingual.text.understanding.and.generation","name":"multilingual text understanding and generation","description":"Processes and generates text in 20+ languages (English, Chinese, French, German, Spanish, Russian, Japanese, Arabic, etc.) through multilingual tokenization and transformer layers trained on diverse language data from The Pile. Supports cross-lingual transfer — knowledge learned in one language can improve performance in others. Enables machine translation, multilingual search, and language-agnostic semantic understanding.","intents":["Build multilingual search and retrieval systems that work across language boundaries","Generate content in multiple languages from single prompts or templates","Perform cross-lingual information retrieval (search in English, retrieve documents in Chinese)","Translate between languages using semantic understanding rather than phrase-based matching"],"best_for":["Global organizations serving multilingual user bases without language-specific model management","Researchers studying cross-lingual transfer and multilingual representation learning","Teams building international applications with unified model infrastructure","Companies needing translation without external API dependencies"],"limitations":["Multilingual performance uneven across languages — English performance ~95% of monolingual baseline, but low-resource languages (Vietnamese, Thai) 60-70% of baseline","Tokenization inefficiency for non-Latin scripts (Chinese, Arabic, Japanese) — 2-3x more tokens per character, reducing effective context window","Cross-lingual transfer limited to related language families — Germanic languages benefit more than distant language pairs","No language identification — model may mix languages or generate in wrong language without explicit prompting","Translation quality lower than specialized translation models — 5-15 BLEU points below Google Translate"],"requires":["GPU with 40GB+ VRAM for inference","Multilingual tokenizer (SentencePiece or BPE with multilingual vocabulary)","PyTorch 1.11+","Python 3.8+","Optional: language identification library (langdetect, fasttext) for preprocessing"],"input_types":["text in any supported language","language-tagged prompts (for explicit language control)","code-switched text (mixing multiple languages)"],"output_types":["generated text in target language","cross-lingual embeddings (for multilingual search)","language-specific probabilities (for language identification)"],"categories":["text-generation-language","memory-knowledge"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox__cap_8","uri":"capability://memory.knowledge.long.context.reasoning.with.retrieval.augmentation","name":"long-context reasoning with retrieval augmentation","description":"Extends effective context window beyond 2048 tokens through retrieval-augmented generation (RAG) — retrieving relevant documents from external knowledge bases and conditioning generation on retrieved context. Implements dense passage retrieval using embeddings to find relevant documents, then feeds top-k documents as context to the language model for generation. Enables reasoning over large document collections without fine-tuning.","intents":["Answer questions over large document collections by retrieving relevant context dynamically","Generate summaries of long documents by retrieving key sections and synthesizing","Implement fact-checking by retrieving supporting evidence and verifying claims","Build knowledge-grounded dialogue systems that cite sources"],"best_for":["Teams building question-answering systems over proprietary document collections","Organizations implementing fact-checking or evidence-based content generation","Researchers studying retrieval-augmented generation and knowledge grounding","Companies building knowledge-grounded chatbots with source attribution"],"limitations":["Retrieval quality critical — poor retrieval (wrong documents) leads to hallucination and incorrect answers; no mechanism to detect retrieval failures","Requires external vector database and retrieval infrastructure — adds complexity and latency (50-200ms per query for retrieval)","Retrieved context must fit in 2048-token window — limits to 5-10 typical documents, may miss relevant information","No joint optimization of retrieval and generation — retriever and generator trained separately, suboptimal end-to-end performance","Hallucination still occurs even with retrieved context — model may ignore or contradict retrieved documents"],"requires":["Vector database (Pinecone, Weaviate, Milvus, Qdrant) with indexed documents","Embedding model for dense passage retrieval (can use GPT-NeoX embeddings or specialized retriever)","GPU with 40GB+ VRAM for generation","PyTorch 1.11+","Python 3.8+","Document collection indexed and preprocessed (chunked into passages)"],"input_types":["user queries (natural language questions)","document collection (text, PDFs, web pages)","retrieval parameters (top-k, similarity threshold)"],"output_types":["generated answers with retrieved context","source citations (document IDs, passage indices)","retrieval scores (confidence in retrieved documents)"],"categories":["memory-knowledge","search-retrieval","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0}],"trust":{"score":21,"verified":false,"data_access_risk":"high","permissions":["GPU with minimum 40GB VRAM (A100 80GB, H100, or equivalent) for full precision","PyTorch 1.11+ or compatible deep learning framework","Python 3.8+","825GB disk space for full model weights (or 5-10GB for quantized versions)","CUDA 11.3+ for GPU acceleration (or CPU inference with 10-50x latency penalty)","Labeled instruction dataset (minimum 500-1000 examples for meaningful adaptation)","GPU with 40GB+ VRAM for full fine-tuning, or 24GB+ for LoRA-based adaptation","PyTorch 1.11+ with distributed training support (torch.nn.parallel.DistributedDataParallel)","Fine-tuning framework (HuggingFace Transformers, DeepSpeed, or Megatron-LM)","2-8 GPUs with 40GB+ VRAM each (A100, H100, or equivalent)"],"failure_modes":["20B parameters requires 40GB+ VRAM for full precision inference (16-bit requires 40GB, 8-bit quantization reduces to ~20GB but adds latency)","Inference speed significantly slower than optimized commercial APIs — ~50-100ms per token on A100 vs 10-20ms for GPT-3.5","Knowledge cutoff at April 2022; cannot access real-time information without external retrieval","No instruction-tuning or RLHF applied in base model — requires additional fine-tuning for chat/instruction-following tasks","Context window limited to 2048 tokens, insufficient for long-document analysis without chunking","Base model lacks instruction-tuning — raw outputs often verbose, unfocused, or off-topic without fine-tuning","Fine-tuning requires 10-100GB GPU memory depending on batch size and sequence length","No built-in safety alignment (RLHF) — fine-tuned models inherit base model's lack of refusal behaviors","Instruction-following quality depends heavily on fine-tuning dataset quality and size; small datasets (<1K examples) often underperform","LoRA adaptation adds inference latency (~5-10%) due to rank decomposition overhead","builder identity is not verified yet","no observed match outcomes yet"],"rank_breakdown":{"adoption":0.05,"quality":0.33,"ecosystem":0.25,"match_graph":0.25,"freshness":0.5,"weights":{"adoption":0.35,"quality":0.2,"ecosystem":0.1,"match_graph":0.3,"freshness":0.05}},"observed_outcomes":{"matches":0,"success_rate":0,"avg_confidence":0,"top_intents":[],"last_matched_at":null},"maintenance":{"status":"inactive","updated_at":"2026-06-17T09:51:03.041Z","last_scraped_at":"2026-05-03T14:00:27.894Z","last_commit":null},"community":{"stars":null,"forks":null,"weekly_downloads":null,"model_downloads":null,"model_likes":null}},"distribution":{"claim_url":"https://unfragile.ai/submit?claim=gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox","compare_url":"https://unfragile.ai/compare?artifact=gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox"}},"signature":"ZmNHK2dBPzloW+HRJrx7MWIqLRDSJuTjfEopEFfgc4tcfC7escIHUye+FOSwOwooQr9UU1bmhx/ochcbEd95BQ==","signedAt":"2026-06-20T13:49:39.158Z","signedBy":"unfragile.ai","version":1},"_links":{"self":"https://unfragile.ai/api/v1/passport/gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox","artifact":"https://unfragile.ai/gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox","verify":"https://unfragile.ai/api/v1/verify?slug=gpt-neox-20b-an-open-source-autoregressive-language-model-gpt-neox","publicKey":"https://unfragile.ai/api/v1/trust-passport-public-key","spec":"https://unfragile.ai/trust","schema":"https://unfragile.ai/schema.json","docs":"https://unfragile.ai/docs"}}