license: apache-2.0
base_model: mistralai/Mistral-7B-Instruct-v0.3
tags:
- rys
- layer-duplication
- reasoning-circuits
- gguf
- sovereign-collection-v2
Mistral-7B-v0.3-RYS-18-23
Mistral-7B-Instruct-v0.3 with layers 18-22 duplicated. The mid-late stack reasoning circuit — the same depth-position as the Llama-3.1-8B reasoning circuit — runs twice on every forward pass.
32 base layers → 37 after duplication. No training, no merging, no weight changes.
Reasoning 41.18% → 58.83% (+17.65). Math 0.534 → 0.5856 (+5.16). EQ 88.52 → 87.19 (−1.33).
Results
| Metric | Baseline | RYS (18,23) | Delta |
|---|---|---|---|
| Math | 0.534 | 0.5856 | +5.16 |
| EQ | 88.52 | 87.19 | −1.33 |
| Reasoning | 41.18% | 58.83% | +17.65 |
The mid-baseline reasoning lift. Mistral-7B-v0.3 places its primary reasoning circuit at the same depth-fraction as Llama-3.1-8B (both peak at layers ~18-22 of a 32-layer stack, block-size 4-5). The position is shared across architectures. The magnitude differs: Mistral's weaker reasoning baseline (41.18% vs Llama's 82.35%) gives it more recoverable headroom, and 28 of 66 configurations boost reasoning >5% (vs Llama-3.1-8B's 15 boosters). Position is architecture-determined; magnitude is baseline-determined.
Pick this when you want strong reasoning out of a 7B class model with minimal EQ trade-off.
Usage
llama-server -m Mistral-7B-v0.3-RYS-18-23-Q4_K_M.gguf -ngl 99
Full sweep data
66 configurations tested. (18,23) block-5 is the best-combined pick. Full per-config sweep + cross-architecture analysis: v2 dataset.
Part of the RYS Sovereign Collection v2.
Where this sits in the Sovereign Collection
v1 — Qwen2.5 cross-scale + Qwen3-32B headline crossover. 5 model repos.
v2 — cross-architecture corpus. 21 model variants across 10 architecture families. Inverse correlation (r = −0.726): weak baselines lift more, in their weakest dimension. Mistral's matched-position / different-magnitude relationship with Llama-3.1-8B is one of the cleanest demonstrations of the position-vs-magnitude factorization. 13 deployable RYS-applied weight repos covering every non-zero-lift variant.
Cross-architecture comparator: john-broadway/Llama-3.1-8B-RYS-18-22-GGUF — same circuit-position, stronger baseline, modest lift.
Credit
John Broadway, with collaboration from Claude (Opus 4.6 in April 2026 sweep generation and build pipeline; Opus 4.7 in May 2026 cross-architecture analysis and publication). Original RYS method by David Ng on Qwen2-72B; sweep + probe toolkit by alainnothere.