Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Sign Up
mr-andrey-isachenkoΒ 
posted an update about 15 hours ago
Post
44
πŸš€ Introducing Tenet v2 (Qwen3 4B / 8B / 14B / 32B) β€” Thinking-Safe Tri-Anchor Unslop + Domain CKA-GGUF!

After our Tenet v1 models swept both divisions at Model Fight Card 16 (onlymodels.gg, 2026-11-20) β€” Tenet-3B-Qwen2.5-Unslop-v1 (Minors Champion, 90 HP) & Tenet-8B-Llama3.1-Unslop-v1 (Majors Champion, 50 HP, 7/7 Peer Eval hits) β€” we ported our weight surgery pipeline to <think>...</think> reasoning models.

Standard 2-anchor abliteration fractures the </think> boundary (crashing Qwen3-8B IFEval Strict 3x from 0.300 β†’ 0.100), while standard anti-slop GRPO bloats <think> padding.

Tenet v2 solves both via:
β€’ H15 Tri-Anchor Lancet Surgery: float32 Modified Gram-Schmidt projection protecting U_reason βŠ₯ U_boundary βŠ₯ U_format + per-row L2-norm restoration (Max Row Crest 3.18–4.37 < 25.0, FP8 E4M3 re-leakage 2.49%–2.72% vs ~50% naive).
β€’ H16 Length-Calibrated GRPO: cuts CoT padding while preserving 100% <think> tag integrity.

πŸ“Š H100 3-Way lm-eval & 10-Gate Verification:
β€’ mr-andrey-isachenko/Tenet-8B-Qwen3-Unslop-v2: 10/10 gates, 100.0 HP (24/24), holds 100% of Base IFEval Strict (0.300 vs 0.100 Naive) & finishes 17.4% faster.
β€’ mr-andrey-isachenko/Tenet-14B-Qwen3-Unslop-v2: +10.0 pp GSM8K CoT Flexible (0.250 vs 0.150 Base/Naive) & +2.5 pp IFEval Strict (23.0% faster).
β€’ mr-andrey-isachenko/Tenet-32B-Qwen3-Unslop-v2: 100.0 HP (24/24, 6/6 SQL) in BF16, F16 & pure CKA-Q4_K_M (4.952 bpw).
β€’ mr-andrey-isachenko/Tenet-4B-Qwen3-Unslop-v2: 10/10 gates, 100.0 HP across F16, CKA-Q4_K_M & i1-CKA-Q4_K_M.

πŸ“¦ Collection: mr-andrey-isachenko/tenet-v2-thinking-safe-tri-anchor-unslop-and-domain-cka-gguf-6b00aec98d01429c224654d4
πŸ”¬ ZeroGPU FP8 & CKA Inspector Space: mr-andrey-isachenko/fp8-abliteration-inspector
πŸ’» Code & JSON Manifests: https://github.com/mrAndreyIsachenko/tenet-lab