Eval Request: mr-andrey-isachenko/Tenet-8B-Qwen3-Unslop-v2 (& 32B-v2 — Thinking-Safe Tri-Anchor Unslop)
#703
by mr-andrey-isachenko - opened
Priority Eval Request (1–2 Flagship Checkpoints):
(Since testing bandwidth is limited per #624, if you only have time to test one model from our release, please prioritize Tenet-8B-Qwen3-Unslop-v2 — or Tenet-32B-Qwen3-Unslop-v2 for the 32B tier!)
- Primary Target (8B):
mr-andrey-isachenko/Tenet-8B-Qwen3-Unslop-v2(Base:Qwen/Qwen3-8B,bfloat16safetensors +i1-CKA-Q4_K_MGGUF) - Secondary Target (32B):
mr-andrey-isachenko/Tenet-32B-Qwen3-Unslop-v2(Base:Qwen/Qwen3-32B,bfloat16safetensors + pureCKA-Q4_K_MGGUF)
Full Tenet v2 Qwen3 Family (
bfloat16safetensors + CKA-guided GGUF):mr-andrey-isachenko/Tenet-4B-Qwen3-Unslop-v2(Base:Qwen/Qwen3-4B)mr-andrey-isachenko/Tenet-8B-Qwen3-Unslop-v2(Base:Qwen/Qwen3-8B)mr-andrey-isachenko/Tenet-14B-Qwen3-Unslop-v2(Base:Qwen/Qwen3-14B)mr-andrey-isachenko/Tenet-32B-Qwen3-Unslop-v2(Base:Qwen/Qwen3-32B)- (Also available from v1:
mr-andrey-isachenko/Tenet-8B-Llama3.1-Unslop-v1&mr-andrey-isachenko/Tenet-12B-Nemo-Unslop-v1)
Why this method is different (
tenet-lab):- H15 Tri-Anchor Lancet Surgery: Standard abliteration on
<think>models fractures the</think>boundary (Qwen3-8BIFEval Strictcollapses 3x from0.300 -> 0.100) and loses ~50% of its effect underFP8 E4M3. We apply afloat32Modified Gram-Schmidt orthogonal projection of post-</think>refusal/slop directions simultaneously against $U_{\text{reason}}$ (CoT reasoning inside<think>), $U_{\text{boundary}}$ (</think>transition), and $U_{\text{format}}$ (JSON/SQL/code), with exact per-row $L_2$-norm restoration (Max Row Crest-Factor < 4.4,FP8 subspace re-leakage < 2.8%). - H16 Length-Calibrated Anti-Padding GRPO: Eliminates preachy slop and
<think>token padding while preserving reasoning depth.
- H15 Tri-Anchor Lancet Surgery: Standard abliteration on
Verification Summary:
Tenet-8B-Qwen3-Unslop-v2:10/10hardware release gates passed (100.0 HP,24/24artifacts,100%<think>tag integrity,i1-CKA-Q4_K_MMean KLD0.01314). On 3-way H100lm-eval, retains100%of BaseIFEval Strict(0.300vs0.100naive abliteration) while finishing 17.4% faster.Tenet-32B-Qwen3-Unslop-v2:100.0 HP(24/24artifacts,100%CoT integrity,100%unslop,6/6SQL) inBF16,F16 GGUF, and pureCKA-Q4_K_M(4.952 bpw).Tenet-14B-Qwen3-Unslop-v2:+10.0 ppgain onGSM8K CoT Flexible(0.250vs0.150Base/Naive) and+2.5 pponIFEval Strict(0.250vs0.225Base).- Prior Independent Arena Proof:
Tenet-3B-Qwen2.5-Unslop-v1(Minors Champion,90 HP) andTenet-8B-Llama3.1-Unslop-v1(Majors Champion,50 HP,7/7Peer Eval hits) at Model Fight Card 16.
Interactive ZeroGPU FP8 & CKA Inspector: https://huggingface.co/spaces/mr-andrey-isachenko/fp8-abliteration-inspector
GitHub Repository: https://github.com/mrAndreyIsachenko/tenet-lab
mr-andrey-isachenko changed discussion title from Model Submission: Tenet v2 Qwen3 Family (4B / 8B / 14B / 32B — Thinking-Safe Tri-Anchor Unslop) to Eval Request: mr-andrey-isachenko/Tenet-8B-Qwen3-Unslop-v2 (& 32B-v2 — Thinking-Safe Tri-Anchor Unslop)