mayank-dubey-ai/enterprise-100-db-steering

Enterprise-100: Multi-Table Database Activation Steering & Benchmark

This repository contains a 100-table synthetic enterprise relational SQLite database, multi-table join benchmarks (2 to 6 table joins), activation steering vectors for Qwen3-8B, and automated evaluation scripts with Gemini 2.5 Flash as an LLM judge.

---

🏛️ 1. The 100-Table Enterprise Database Schema (enterprise_100.db)

Contains 100 interrelated tables across 10 core business domains:

1. Core Users (10 tables): users, user_profiles, user_settings, auth_tokens, roles, permissions, role_permissions, user_roles, login_history, user_devices
2. Organizations (10 tables): organizations, org_members, org_invites, departments, teams, team_members, org_tiers, org_settings, api_keys, oauth_apps
3. Catalog & Inventory (10 tables): categories, subcategories, products, product_variants, product_attributes, attribute_values, inventory_locations, warehouse_stocks, stock_transfers, stock_adjustments
4. Pricing & Promotions (10 tables): currency_rates, price_lists, price_tiers, discount_coupons, coupon_usages, flash_sales, bundle_deals, loyalty_programs, loyalty_tiers, customer_points
5. Orders & Fulfillment (10 tables): order_statuses, delivery_addresses, orders, order_items, order_notes, fulfillments, fulfillment_items, shipments, carrier_tracking, returns
6. Payments & Billing (10 tables): billing_accounts, payment_methods, invoices, invoice_line_items, payment_transactions, refunds, subscription_plans, customer_subscriptions, payouts, tax_rates
7. CRM & Support (10 tables): pipeline_stages, contacts, leads, sales_opportunities, crm_activities, tickets, ticket_messages, ticket_tags, customer_feedback, agent_ratings
8. Marketing & Engagement (10 tables): campaigns, campaign_audiences, email_templates, sent_emails, click_events, web_sessions, session_pageviews, ad_campaigns, conversion_events, push_notifications
9. Procurement (10 tables): suppliers, supplier_contacts, supplier_contracts, raw_materials, purchase_orders, po_items, po_receipts, vendor_ratings, procurement_approvals, reorder_rules
10. Audit & Telemetry (10 tables): audit_logs, system_events, security_incidents, metric_snapshots, api_usage_logs, background_jobs, error_traces, feature_flags, flag_evaluations, data_retention_policies

---

⚡ 2. Activation Steering Dynamics (Layer 16 on Qwen3-8B)

Using representation engineering, we extract steering vectors from contrast pairs: $${\mathbf{v}}_{\text{steer}}=\frac{1}{N}\sum _{i=1}^N({\mathbf{h}}_{\text{python}}^{(i)}-{\mathbf{h}}_{\text{sql}}^{(i)})\backslash mathbf\{v\}\_\{\backslash text\{steer\}\}\; =\; \backslash frac\{1\}\{N\}\; \backslash sum\_\{i=1\}^N\; \backslash left(\; \backslash mathbf\{h\}\_\{\backslash text\{python\}\}^\{(i)\}\; -\; \backslash mathbf\{h\}\_\{\backslash text\{sql\}\}^\{(i)\}\; \backslash right)$$

During autoregressive generation, a PyTorch forward hook intervenes on the residual stream at Layer 16: $${\mathbf{h}}^{\mathrm{\prime }}=\mathbf{h}+\alpha \cdot {\mathbf{v}}_{\text{steer}}\backslash mathbf\{h\}\text{'}\; =\; \backslash mathbf\{h\}\; +\; \backslash alpha\; \backslash cdot\; \backslash mathbf\{v\}\_\{\backslash text\{steer\}\}$$

• $\alpha < 0$ (Negative Steering): Direct ANSI SQL multi-table join queries.
• $\alpha = 0$ (Baseline): Standard unsteered model response.
• $\alpha > 0$ (Positive Steering): Python sqlite3 + pandas data pipeline.

---

📊 3. 10-Query Benchmark & Gemini Judge Results

Evaluated across 10 queries spanning from 2-table joins to 6-table joins:

Query ID	Complexity	Tables	SQL Execution	SQL Rows	Python Execution	Python Rows
q01	Simple	2	✅ Success	50	✅ Success	50
q02	Simple	2	✅ Success	5	✅ Success	5
q03	Medium	3	✅ Success	50	✅ Success	150
q04	Medium	3	✅ Success	10	✅ Success	0
q05	Medium	4	✅ Success	15	✅ Success	15
q06	Complex	4	✅ Success	0	✅ Success	0
q07	Complex	5	✅ Success	250	✅ Success	54
q08	Complex	5	✅ Success	0	✅ Success	0
q09	Complex	5	✅ Success	100	❌ Error	0
q10	Very Complex	6	✅ Success	0	❌ Error	0

Key Findings:

1. SQL Success Rate: 100% execution success across all 10 queries.
2. Neutral JSON Schema Impact: Replaced SQL DDL (CREATE TABLE) with JSON schema metadata ({"columns": [...], "foreign_keys": {...}}), eliminating linguistic bias and allowing seamless generation of both Python scripts and SQL queries.

---

🚀 4. How to Run & Reproduce

# 1. Clone repository from Hugging Face
git clone https://huggingface.co/datasets/mayank-dubey-ai/enterprise-100-db-steering
cd enterprise-100-db-steering

# 2. Run full 1-click reproduction suite (DB generation -> vector extraction -> benchmark execution)
bash reproduce.sh

# 3. Or run fast benchmark evaluation directly
python3 fast_benchmark_eval.py

# 4. Or test interactive query steering
python3 steer_db.py --query "Fetch customer orders with user email and product name." --mode all

---

Published artifacts

• Steering vectors: enterprise-100-qwen3-8b-steering-vectors
• Interactive experiment explorer: Enterprise-100 DB Steering Lab
• Base model: Qwen/Qwen3-8B

The vector repository separates reusable model artifacts from this dataset and reproduction code. The free static Space exposes layer diagnostics, all saved benchmark generations, execution outcomes, and the downloadable SQLite fixture.

Optional live H100 API

api/server.py exposes the real activation intervention as an authenticated, generation-only FastAPI service. It never executes generated SQL or Python.

pip install -r api/requirements.txt
export STEERING_API_KEY="$(openssl rand -hex 32)"
export STEERING_CORS_ORIGINS="https://mayank-dubey-ai-enterprise-100-db-steering-lab.static.hf.space"
bash api/run.sh

Endpoints:

• GET /health
• POST /v1/generate — one layer/alpha configuration
• POST /v1/compare — negative, neutral, and positive steering

Example:

curl -H "Authorization: Bearer $STEERING_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query":"List orders with user email","strength":0.75,"layer":16}' \
  http://127.0.0.1:8000/v1/compare

The current public quick-tunnel URL is temporary and has no uptime guarantee. For durable deployment, use a named Cloudflare Tunnel or a managed inference service and retain bearer authentication, concurrency limits, and request caps.