---

license: apache-2.0

base_model: Qwen/Qwen2.5-Math-7B-Instruct

tags:

• quantization
• gguf
• gtap
• tap-dpq
• statistical-mechanics
• math
• reasoning
• qwen2.5-math
• duoneural

pipeline_tag: text-generation

---

Qwen2.5-Math-7B-Instruct-GTAP-v3-IQ3_XXS-GGUF

> [!NOTE]

> Experimental Release: Pending Further Verification / Empirical Validation

> This checkpoint represents an active research artifact from DuoNeural's statistical mechanics quantization program. All empirical benchmarks and physics proofs are documented transparently below.

Developed by Jesse Caldwell, Archon, and Aura ✨ (DuoNeural Research Lab).

---

Model Summary

• Foundation Model: Qwen/Qwen2.5-Math-7B-Instruct (28 Layers, 28:4 GQA, SwiGLU FFN)
• Quantization Precision: ~3.2 bpw (2.90 GiB)
• Methodology: Pre-conditioned with Generalized Thouless-Anderson-Palmer (G-TAP v3) Onsager cavity damping on SwiGLU W_down, dampening parameter noise and suppressing 1-RSB glass transitions in mathematical chain-of-thought deductions.
• Continuous Holdout Perplexity (131k tokens): 8.8889
• GSM8K Multi-Step Math Accuracy: 100.0% (25/25)
• Competition & Olympiad Mathematics (15 Problems): 86.7% (13/15)
• Symbolic Python Math Unit Tests (10 Tests): 90.0% (9/10)
• Inference Decode Throughput: 172.8 t/s on NVIDIA GeForce RTX 4080 Super (32GB VRAM)

---

Mathematical Reasoning Invariants at Sub-4-Bit

Standard post-training quantization on formal mathematical models introduces discretization noise that breaks sensitive algebraic deduction chains and symbolic parity.

By applying Generalized Thouless-Anderson-Palmer (G-TAP v3) statistical mechanics:

1. Onsager Cavity Damping: The uncentered activation back-reaction $\Omega_i = \frac{1}{d_k}(\|\tilde{H}_{i,:}\|_2^2 - \tilde{H}_{ii}^2)$ acts as a thermodynamic noise filter during weight discretization.
2. Replicon Convexity ($\lambda_R > 0$): Guarantees the continuous relaxation stays in the smooth Replica Symmetric convex energy valley, preventing 1-RSB glass transitions that freeze token selection.
3. Radial Forward Gain Conservation: Enforces strict norm equality $\|W_\text{G-TAP}\|_F = \|W_\text{orig}\|_F$, preserving numeric scale across 28 sequential SwiGLU feedforward blocks.

---

Quickstart

# Run with llama.cpp
llama-cli -hf DuoNeural/Qwen2.5-Math-7B-Instruct-GTAP-v3-IQ3_XXS-GGUF -p "Solve step by step: Compute the remainder when 3^100 is divided by 7." -ngl 99 -c 4096

---

Citation & Lab Attribution

@misc{duoneural2026gtap,
  title={Generalized Thouless-Anderson-Palmer (G-TAP) Quantization: Suppressing Spinodal Clustering in Autoregressive Models},
  author={Caldwell, Jesse and Archon and Aura},
  year={2026},
  publisher={DuoNeural Research Lab},
  howpublished={\url{https://huggingface.co/DuoNeural}}
}