oowm-merge-v1 โ€” TIES proof-of-pipeline merge (SOVOS)

494M param Qwen2-family model. Produced by mergekit TIES self-merge on RunPod (RTX 3090), signed provenance card attached.

Honest status (measurement, not certification)

  • method: TIES self-merge (0.65/0.35, density 0.7) of the 958M refusal-merged sovereign model.
  • what it proves: the mergekit pipeline runs end-to-end and preserves the model. NOT a capability gain โ€” a self-merge is behaviourally identical to its parent by construction.
  • measured: 6/8 on an 8-item governance/refusal/instruction/agreement battery โ€” identical to parent, as expected. n=8 << usable_n=30, so proof-of-pipeline only, not a quotable score.
  • recipe_sha256: c67b5e9fc756e435fc4a8c1e254c3b7f1c5fcfaed29503540ba26826d9e23d39
  • weight_sha256: 65b60a669feb9c814d4a5fc40c9278cc0b59a3e8dcfd7bd1aa3275378db0e6f4
  • signature: Ed25519, see oowm_MERGE.signed.json.

A real capability-gain merge requires two different same-base adapters (SOV4-T / SOV4-OWEM); those are not co-located with this proof and are the next step.

Downloads last month
39
Safetensors
Model size
0.5B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for csoai/oowm-merge-v1

Finetuned
(938)
this model