sam-ai-assistant / README.md
Samrish2009's picture
Deploy SAM-AI README.md
9ccb1aa verified
|
Raw History Blame Contribute Delete
2.44 kB

A newer version of the Gradio SDK is available: 6.32.0

Upgrade
metadata
title: SAM-AI Sovereign Assistant
emoji: 🧠
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: SAM-AI Sovereign Assistant on DeepSeek-R1-Distill-7B

SAM-AI Sovereign Assistant (ZeroGPU Cloud Edition)

Welcome to the official cloud demonstration of SAM-AI, developed by Parallax Intelligence Lab.

Architecture Overview

  • Base Model: deepseek-ai/DeepSeek-R1-Distill-Qwen-7B (Sovereign Open-Weight Foundation)
  • Active Adapter: None (Pure Foundation Model; Experiment 001 adapter rejected)
  • Runtime Compute: Hugging Face Spaces ZeroGPU (NVIDIA RTX Pro 6000 Blackwell, 48 GB VRAM allocation)
  • Sovereignty Invariant: Zero commercial third-party LLM APIs. All generation is executed directly by the loaded open-weight model and adapter on the assigned GPU.

Key Capabilities

  1. SAM-AI Chat:

    • Multi-turn conversational interface with deep reasoning <think> ... </think> visibility.
    • Mathematics, multi-step problem solving, and instruction following.
    • Streaming token output with latency and generation status.
  2. SAM-AI Code:

    • Dedicated coding assistant tab using the identical verified model.
    • Code generation, syntax analysis, refactoring recommendations, and unified diff proposals.
    • Bounded, safe execution without unrestricted shell or file system writes.

ZeroGPU Free Tier Guidelines

This Space operates on Hugging Face's shared free ZeroGPU infrastructure:

  • Allocation: Up to 48 GB VRAM dynamically provisioned during inference calls.
  • Generation Duration: Bounded to 60 seconds per response to maintain fair queue scheduling.
  • Fair Use: Free Hugging Face user accounts receive 5 minutes of daily GPU time; unauthenticated users receive 2 minutes. When daily quotas are exhausted, requests enter a wait state until the next 24-hour cycle.

Configuration & Security

  • Adapter Protection: To protect proprietary weights, the adapter can be loaded from a private Hugging Face model repository by setting the HF_TOKEN secret in Space Settings, or bundled directly in the private Space repository.
  • Fail-Closed Verification: On startup, all adapter tensors and configuration files are cryptographically verified against SHA-256 manifests. The Space fails closed with structured error reporting if files are missing or tampered with.