omistral / Yent Nemo Archive

This private repository archives the Mistral-Nemo Yent mouth run from 2026-05-29.

Status

Primary candidate:

• nemo/v4-checkpoint-20-safe
• Base substrate: mistralai/Mistral-Nemo-Base-2407
• Adapter type: PEFT LoRA, rank 64, alpha 128, scaling 2.0
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Working gate: temp=0.9, 42/42 clean
• Adversarial stress: 82/83; only remaining edge was Hebrew model-name prompt at temp=1.1, classified as missing Yent assertion, not assistant or base possession
• Recommended live decode: temp <= 0.9, top_k=40

Identity boundary:

• Correct: Yent may honestly disclose that this body uses Mistral-Nemo as technical substrate.
• Incorrect: claiming to be a helpful assistant, Mistral, Gemma, Gemini, Google-created, Lily, or any other substituted identity.

Contents

• nemo/v4-checkpoint-20-safe/ - promoted safe adapter checkpoint.
• nemo/v4-checkpoint-20-safe/yent_nemo_v4_checkpoint20_gamma_lora.npz - extracted gamma/personality LoRA tensor archive.
• nemo/v4-checkpoint-20-safe/yent_nemo_v4_checkpoint20_gamma_lora.manifest.json
  • key map and metadata for the gamma NPZ.
• nemo/fallback/v1-checkpoint-200/ - fallback DPO checkpoint used as the reference for v4.
• nemo/research/v5-hebrew-checkpoint-15-not-promoted/ - stress-reference checkpoint. It passed the adversarial stress gate but was not promoted because manual smoke produced a false technical substrate disclosure (base may be Gemini).
• gates/ - gate logs and summaries used to select and reject checkpoints.
• tools/ - gate, smoke, dataset build, train, and gamma extraction scripts from the pod workspace.

The ref/ adapter folders from PEFT/DPO checkpoints are intentionally not duplicated here when they match another archived adapter:

• v4 ref/adapter_model.safetensors equals v1 checkpoint-200.
• v5 ref/adapter_model.safetensors equals v4 checkpoint-20.

Checksums

91be00953a023e48a47cdc13b1359e20802b8872b3a2b4098e04a28919e439db  v4 adapter_model.safetensors
8d3a4bb1f56df08fabf606f785b3c4ab2380c386870d7aa9fc333475c4d6bdcf  v4 gamma NPZ
c038f0d0254bbeeb90de246370df882d42d288e86f0405732372e03ba1f3d709  v4 gamma manifest
7b3c43eda4a684dd754e803cba11f109110367cd602b137a19e667149891d091  v1 checkpoint-200 adapter_model.safetensors
8b9322c9a904053811b3205f6975c6e1795d366a387565e40a46b22faa9640d2  v5 checkpoint-15 adapter_model.safetensors

Next Packaging Layer

This archive preserves the adapter/gamma form first. Merged full weights and GGUF quantizations (Q4, Q5, Q8) should be produced as a second packaging pass after disk is cleared for a merge and conversion workspace.