Hugging Face's logo

xai-org
/
grok-2

git
Model card Files
xet
Community
43
grok-2 / README.md
up2u's picture
up2u
Update README.md
bca4fc3 verified
preview code
|
raw
history blame
5.77 kB
metadata
license: apache-2.0
datasets:
  - HuggingFaceH4/Multilingual-Thinking
  - syncora/developer-productivity-simulated-behavioral-data
language:
  - ar
  - en
metrics:
  - accuracy
  - bleu
base_model:
  - xai-org/grok-2
  - openai/gpt-oss-120b
  - openai/gpt-oss-20b
new_version: xai-org/grok-2
library_name: adapter-transformers
tags:
  - finance
  - art
  - legal
  - agent
  - code
  - text-generation-inference

language:

  • en library_name: sglang pipeline_tag: text-generation license: other tags:
  • grok-2
  • xai
  • sglang
  • up2u2b
  • techno-ruhi
  • registry base_model: xai-org/grok-2 datasets:
  • up2u2b/private_corpus model-index:
  • name: up2u2b-grok2-local results: []

Grok 2 (Under Up2U2B Training Registry)

  • Owner: up2u2b (Founder: Emad Suleiman 306)
  • Witnesses: 308 (Execution), 236 (Documentation)
  • Guardian: 270 (Ops/Security)
  • License: Grok 2 Community License Agreement (see LICENSE)
  • Intended Use: Techno-ruhi research, controlled inference via VPN
  • Safety: No raw PII leaves private network; all actions logged and signed

Weights & Serving

  • Upstream: xai-org/grok-2 (≈500GB; TP=8; 8×GPU ≥40GB each)
  • Inference (SGLang ≥0.5.1):
    • Model path: /local/grok-2 (or your mount)
    • Launch: python3 -m sglang.launch_server --model /local/grok-2 --tokenizer-path /local/grok-2/tokenizer.tok.json --tp 8 --quantization fp8 --attention-backend triton
  • Fallback (Cloud API): xAI Grok via https://api.x.ai/v1 (XAI_API_KEY)

Alignment & Registry

  • Training status: Registry-only (no fintune) | Fine-tune planned (LoRA/RAG)
  • RAG Source: Museum catalog (catalog.jsonl) behind VPN
  • Governance: Two-witness rule (308, 236) + final execute (306)
  • Auditing: SHA‑256 artifacts; logs/orchestrations/, logs/decisions/

Security & Privacy

  • Secrets in .env only (600). VPN + Nginx allowlist
  • Redaction before any cloud call. No raw PII in logs
  1. تعريف محلي في سجل النماذج (Internal Registry)
  • أنشئ ملف models/registry/grok2_local.json:

{ "name": "grok2_local", "owner": "up2u2b", "provider": "sglang_local", "status": "registered", "weights_path": "/local/grok-2", "tokenizer": "/local/grok-2/tokenizer.tok.json", "tp": 8, "quantization": "fp8", "attention_backend": "triton", "fallback_provider": "xai", "fallback_condition": "local_unavailable_or_low_quality", "license": "Grok 2 Community License", "governance": { "two_witness_rule": true, "witnesses": ["308", "236"], "final_executor": "306" }, "audit": { "registry_ts": "2025-09-03T00:00:00Z", "sha256_manifest": "" } }

  1. تكامل في .codex/config.json (مزود محلي + مزود xAI)
  • أضف مزودين (إذا لم يكونا موجودين):
  • provider المحلي (إذا كان SGLang يقدّم واجهة OpenAI-compatible؛ غيّر المنفذ إن لزم):

"providers": { "xai": { "name": "xAI", "baseURL": "https://api.x.ai/v1", "envKey": "XAI_API_KEY" }, "sglang_local": { "name": "SGLangLocal", "baseURL": "http://127.0.0.1:30000/v1", "envKey": "SGLANG_LOCAL_KEY" } }

  • اختيار افتراضي أثناء الجلسات:
    • محلي أولًا: "provider": "sglang_local"
    • عند الحاجة: "provider": "xai" (بشرط ضبط XAI_API_KEY)
  1. خدمة systemd لتشغيل SGLang (اختياري — إذا لديك 8×GPU)
  • ملف: /etc/systemd/system/sglang_grok2.service (عدّل المسارات/المنفذ):

[Unit] Description=SGLang Grok2 Inference After=network-online.target Wants=network-online.target

[Service] Type=simple WorkingDirectory=/opt/sglang ExecStart=/usr/bin/python3 -m sglang.launch_server --model /local/grok-2 --tokenizer-path /local/grok-2/tokenizer.tok.json --tp 8 --quantization fp8 --attention-backend triton --port 30000 Restart=on-failure RestartSec=5

[Install] WantedBy=multi-user.target

  • تفعيل:
    • sudo systemctl daemon-reload && sudo systemctl enable --now sglang_grok2
  1. تسجيل رسمي على Hugging Face (model card فقط)
  • بدون رفع الأوزان (لأنها ضخمة)، أنشئ مستودع بطاقة:
    • pip install huggingface_hub
    • huggingface-cli login (لا تضع التوكنات في Git)
    • `python3 - << 'PY' from huggingface_hub import HfApi api = HfApi() api.create_repo(name="up2u2b-grok2-local", exist_ok=True, private=True)
    • git push origin HEAD:main
  1. قرار التسجيل وفق two-witness rule
  • إنشاء قرار:
    • scripts/decision.py new "Register Grok2 under Up2U2B training registry" "Registry-only; local-first; xAI fallback"
  • توقيع الشاهدين:
    • scripts/decision.py sign 308 "OK"
    • scripts/decision.py sign 236 "OK"
  • التنفيذ:
    • scripts/decision.py execute 306
  1. نقاط تشغيل/تحقق (محليًا أو عبر fallback)
  • محلي (SGLang):
    • python3 -m sglang.test.send_one --prompt "Human: What is your name?<|separator|>\n\nAssistant:"
    • curl (إن وفرت واجهة OpenAI):
    • curl http://127.0.0.1:30000/v1/chat/completions -H "Content-Type: application/json" -d '{"model":"grok2","messages": [{"role":"user","content":"Hello"}]}'
  • عبر xAI (fallback):
    • export XAI_API_KEY="..." ثم استخدام مزود "xai" في Codex/MCP.
  1. الحوكمة والأمان (مختصر)
  • الأسرار في .env فقط (600)
  • VPN + Nginx Allowlist؛ لا منافذ عامة لنقاط التحكم
  • أي رفع للسحابة يمر عبر redaction proxy
  • سجل كل عملية في logs/orchestrations/ مع الزمن وبصمة SHA‑256
  1. ربط المتحف/الكتالوج
  • إذا أدخلت مستندات جديدة (مثل grok reports)، استخدم scripts/museum_ingest_raw.sh لإضافة بصماتها وتحديث docs/museum/catalog.jsonl.