markjoseph2003/JurisSim-32B-v3
JurisSim-32B-v3 by markjoseph2003 is a 32 billion parameter neuro-symbolic legal auditor based on Qwen3-32B, designed to translate natural language legal clauses into Z3 SMT-LIB formal constraints. This model, optimized for AMD Instinct MI300X accelerators, identifies adversarial loopholes with mathematical certainty. It features a dual-agent swarm with a Qwen3-32B Auditor and a Qwen2.5-7B Skeptic for real-time self-correction, excelling in formal verification of legal texts.
Loading preview...
JurisSim-32B-v3: Neuro-Symbolic Legal Auditor
JurisSim-32B-v3, developed by markjoseph2003, is a specialized 32-billion parameter model built on Qwen/Qwen3-32B and fine-tuned using QLoRA. Its core function is to act as a neuro-symbolic legislative stress-tester, converting natural language legal clauses into Z3 SMT-LIB formal constraints. This allows for the identification of adversarial loopholes with mathematical certainty.
Key Capabilities & Features
- Dual-Agent Swarm Architecture: Employs a multi-agent feedback loop consisting of:
- The Auditor (Qwen3-32B): Translates legalese into symbolic logic.
- The Skeptic (Qwen2.5-7B): A dedicated linter agent that dry-runs generated Z3 code and performs real-time syntax error correction.
- Universal Logic Bridge: Pre-trained to understand abstract logical archetypes such as Thresholds, Temporal Loops, and Circularity.
- Architecture v2.0 Sandbox: Automatically injects physical invariants like budget sums and time-forward flow into the analysis.
- MI300X Optimized: Specifically designed for and optimized on the AMD Instinct MI300X Accelerator, leveraging ROCm 6.2 and 192GB VRAM for high-fidelity reasoning.
- Logical Archetypes Covered: Detects issues like Threshold Splitting, Temporal Paradoxes, Jurisdictional Null-Zones, and Circular Deadlocks.
Training & Performance
The model was fine-tuned on a single MI300X GPU, overcoming PyTorch SDPA ROCm bugs through an "Ultra-Stable" workaround involving native eager attention, gradient_accumulation_steps=8 with per_device_train_batch_size=1, and gradient_checkpointing=True. It achieved a final validation loss of 1.676 and a token prediction accuracy of 62.61% for complex Legal English to Python Z3 logic translation, demonstrating excellent generalization.