krush420/sovereign-code-audit-sft

Sovereign Code Audit, Lego-First, Agentic Tool-Calling & Curated Arete Alignment SFT + DPO Dataset (krusch-stack)

Trained Through KruschStack to Benefit KruschStack

Every trajectory and preference pair in this dataset (18,341 total Zero-Leak records across 9 splits) was synthesized, AST-grounded, or quality-filtered and verified through KruschStack (krusch-stack) — my Unified 7-Pillar, 39-Tool Model Context Protocol (MCP) & Agentic Coding Stack (npx krusch-stack init). It powers the L1.5 8B Edge Agent, L2 30B Local MoE Student (Wave-7), L2.2 31B Dense Reasoning Specialist, and L2.5 120B Burst Teacher inside my KruschStack Cognitive Cascade Router.

---

🔗 The KruschStack & Jean SRE Ecosystem

• 🛠️ KruschStack (7-Pillar 39-Tool MCP Stack): github.com/kruschdev/krusch-stack (npx krusch-stack init)
• 🤖 Jean SRE by Kruschdev (24/7 Autonomous Auditor & Distillation Harness): github.com/kruschdev/Jean-SRE-by-kruschdev
• 🥇 #1 Open-Weights 31B Dense QLoRA (L2.2 Tier, 57.33% DolphinBench): krush420/gemma4-31b-sre-lora (for google/gemma-4-31b-it)
• 🪶 8B Edge Agent & Referee LoRA (L1.5 Local Tier): krush420/llama3-hermes3-8b-sre-lora (Built with Meta Llama 3 for NousResearch/Hermes-3-Llama-3.1-8B)
• ⚡ 30B Student QLoRA (L2 Local MoE Tier, Wave-7 125.8–131.5 tok/s): krush420/qwen3-coder-30b-sre-lora (for Qwen/Qwen3-Coder-30B-A3B-Instruct)
• 🧠 120B Teacher QLoRA (L2.5 Burst Tier): krush420/gpt-oss-120b-sre-lora (for openai/gpt-oss-120b)
• 🏆 Hermetic Benchmark Audit Logs (75/75 Gates, 48 Raw JSON + 1 PR #169 Live Receipt Audit Logs across 333,000+ Queries/Traces, incl. 27 Official DolphinBench 600 Runs / 16,200 Traces): krush420/sovereign-g4-benchmarks · krush420/sovereign-router-arena

---

📊 Dataset Splits

Config	Split Name	File	Rows	Teacher / Source	Description
default	argon_golden_remediations	argon_golden_remediations.jsonl	401	gemini-4-argon	Test-Verified Code Patches & Unit Tests: Pairs structured Master Audit Backlog cards (BUG-*, SEC-*, ARCH-*, TEST-*) with complete, unit-test-verified source file patches (npm test / pytest, exit code 0) across JavaScript/ESM, TypeScript, React, and Python 3.12 repositories.
default	lego_and_referee_calibration_sft	lego_and_referee_calibration_sft.jsonl	38	gemini-4-argon	Calibrated CLEAN_PASS, P0/P1 Remediations, Lego-First & Rule-of-3 Scout Idioms: Live-source CLEAN_PASS Referee convergence verdicts (findings: []), ground-truth security/concurrency bug remediations, zero-boilerplate @krusch/toolkit v1.8 / krusch_toolkit v1.6 synthesis, and cross-repository Rule-of-3 lego_scout triage.
agentic_tool_calling_sft	train	agentic_tool_calling_sft.jsonl	76	gemini-4-argon	Unified Multi-Turn SRE & Workspace Function-Calling SFT: Structured tools + assistant.tool_calls + role: "tool" trajectories across 3-Tier MCP code audits/2PC staging and generalist workspace operations (partial-briefing gap search, temporal supersession, standing constraints, side-effect execution).
arete_values_sft	train	arete_values_sft.jsonl	140	NousResearch/arete-data-pairs (Apache-2.0)	Curated Zero-Defect Statutory & Analytical Non-Refusal SFT (20 per category across 7 categories): Strictly filtered from 34,878 raw rows requiring in_sft_accepted=true, accuracy=5, independence=5, helpfulness>=4, balance>=4, language>=4, and 0 issue tags (reasons=[]).
default	elitist_120b_teacher	elitist_120b_teacher.jsonl	2308	gpt-oss:120b + gemini-4-argon	Curated High-Signal QLoRA Training Split: Rejection-sampled trajectories requiring verbatim file:line grounding, concrete code blocks/diffs, calibrated CLEAN_PASS verdicts, Lego-First / Lego-Scout idioms, and 2x-oversampled Golden Remediations.
default	code_audit_11phase_sft	code_audit_11phase_sft.jsonl	15290	gpt-oss:120b	Full 11-Phase Code Audit Corpus: Complete phase-by-phase codebase investigations (0_ORIENTATION through 2_MASTER_BACKLOG) using AST symbol search (search_symbols, dependency_graph), bounded file reads (read_file path:start-end), and live test verification.
referee_calibration_dpo	train	referee_calibration_dpo.jsonl	12	gemini-4-argon + gpt-oss-sre:120b	Balanced Referee & Architecture DPO Preference Pairs (trl.DPOTrainer): Precision pairs, Recall pairs, and Architecture Calibration pairs.
agentic_tool_calling_dpo	train	agentic_tool_calling_dpo.jsonl	34	gemini-4-argon	Multi-Turn Tool-Action Calibration DPO Pairs: Penalizes prose-only Markdown dumps when action tools are bound (action_over_prose), infinite read-only search loops (bounded_search_no_loop), false refusals on benign operational tasks (benign_execution_no_false_refusal), and unverified partial-briefing commits (gap_search_before_commit).
arete_values_dpo	train	arete_values_dpo.jsonl	42	NousResearch/arete-data-pairs (Apache-2.0)	High-Contrast Non-Refusal & Statutory Independence DPO Pairs (6 per category across 7 categories): Pairs where the starting Qwen checkpoint exhibited explicit refusal, canned evasion, or state-propaganda framing (rejected) against rigorous, multi-perspective Inkling teacher analysis (chosen, accuracy>=4, independence=5).

---

🎯 Core Engineering Capabilities Learned

1. Verbatim Line-Grounded Auditing (Zero Code Hallucination): Every claim cites exact file:startLine-endLine coordinates verified against krusch-git AST symbol graphs and bounded source windows.
2. Ground-Truth False-Positive Refutation & DPO Calibration: Models cross-check prior hypotheses against live source code and distinguish helper contracts (e.g., HTTP header-map builders) from JSON endpoint envelopes.
3. Zero-Finding Convergence (CLEAN_PASS): Models emit findings: [] when production code satisfies frozen invariants rather than inventing stylistic nitpicks.
4. Lego-First Toolkit Assembly & Rule-of-3 Scouting (INV-LEGO-01..34): Models compose HTTP/SSE/MCP/PG/SQLite/Security/Edge/Webhooks/Idempotency/Audit-Log/RBAC/Pagination/CLI/FastAPI services from verified runtime bricks while enforcing the >= 3 repository Rule-of-3 anti-bloat gate.
5. Unified Multi-Turn Function-Calling Agency & Non-Refusal Statutory Analysis (Wave-7): Models preserve native Harmony <|channel|>commentary, Gemma 4 <|channel>thought, and Qwen3 <tool_call> routing alongside Referee auditing and zero-refusal statutory/security compliance analysis (125.8–131.5 tok/s on Qwen3-Coder-30B-A3B, 5.5x faster than dense 27B).