Update README.md

README.md

@@ -1,95 +1,59 @@
-# =========================================
-# VERIDIAN-BETA | NULLXES AI MODEL CARD
-# =========================================
-
-model:
-  name: "VERIDIAN-BETA"
-  family: "VERIDIAN"
-  version: "2B-beta"
-  organization: "NULLXES"
-  license: "proprietary"
-
-identity:
-  description: >
-    VERIDIAN-BETA is a lightweight but capable language model
-    developed as part of the NULLXES Digital Workforce Infrastructure.
-    Designed for real-time reasoning, conversational agents, and
-    embedded AI employees operating under constrained latency.
-
-  philosophy: >
-    Not a chatbot. A node in a distributed cognitive system.
-    Built to integrate, not to entertain.
-
-base_model:
-  name: "VERIDIAN 2B"
-  checkpoint: "step_0004750.pt"
-  architecture: "Transformer (causal LM)"
-  tokenizer: "BPE (131k vocab, Qwen-style)"
-
-training:
-  pipeline:
-    - stage: "continued_pretraining"
-      source: "step_0004750.pt"
-      notes: >
-        Domain adaptation and stabilization on curated corpora.
-
-    - stage: "quick_clean_sft"
-      source: "pretrain_stage2_5 step_0001000.pt"
-      notes: >
-        Lightweight supervised fine-tuning on cleaned instruction data
-        for improved dialogue alignment and response clarity.
-
-  techniques:
-    - mixed_precision: "bf16"
-    - optimizer: "AdamW"
-    - gradient_checkpointing: true
-    - streaming_dataset: true
-
-capabilities:
-  - real-time conversational inference
-  - instruction following
-  - structured response generation
-  - integration with digital avatars
-  - low-latency reasoning for WebRTC pipelines
-
-limitations:
-  - reduced depth vs larger frontier models
-  - sensitive to prompt structure
-  - not designed for long-horizon reasoning (>8k context without extension)
-
-intended_use:
-  - digital employees (HR, support, onboarding)
-  - real-time voice agents (WebRTC)
-  - embedded enterprise copilots
-  - internal NULLXES infrastructure components
-
-files:
-  weights: "model.safetensors"
-  tokenizer: "tokenizer.json"
-  configs: "configs/"
-
-deployment:
-  format: "safetensors"
-  device: "GPU (H100/H200/A100 recommended)"
-  inference_mode: "low-latency / streaming-ready")
-
-security:
-  notes: >
-    Model is distributed without training data.
-    Access is controlled via API endpoints only in production environments.
-    No raw weights exposure in enterprise deployments.
-
-contact:
-  organization: "NULLXES"
-  ceo_email: "ceo@nullxes.com"
-  alias: "MagistrTheOne"
-
-meta:
-  created: "2026-04"
-  status: "beta"
-  internal_codename: "VERIDIAN-BETA-4750-SFT"
-  tagline: "Digital cognition. Deployed."
-
-# =========================================
-# END OF FILE
-# =========================================
+---
+language:
+- en
+- ru
+
+license: other
+
+tags:
+- nullxes
+- veridian
+- llm
+- transformer
+- conversational-ai
+- real-time
+- digital-employees
+
+pipeline_tag: text-generation
+
+library_name: pytorch
+
+base_model: veridian-2b
+
+model-index:
+- name: veridian-beta
+  results: []
+
+datasets:
+- nullxes-internal
+
+metrics: []
+
+widget:
+- text: "Hello. You are VERIDIAN. Respond clearly and precisely."
+
+co2_eq_emissions:
+  emissions: 0.0
+
+---
+# VERIDIAN-BETA
+
+VERIDIAN-BETA is a lightweight language model developed by NULLXES
+for real-time AI workforce infrastructure.
+
+## 🔥 Highlights
+- Low latency inference
+- Designed for WebRTC pipelines
+- Optimized for digital employees
+
+## 📦 Files
+- model.safetensors
+- tokenizer.json
+- configs/
+
+## 🧠 Base Model
+VERIDIAN 2B (step_0004750.pt)
+
+## 📬 Contact
+ceo@nullxes.com
+MagistrTheOne