15 days ago
-
Model card: held-out selective risk per set for the frozen E36 policy (5 % budget missed on OOD extremes and routing task)
17 days ago
-
remove The name section
-
card: add the name, use cases with bounds, E31 correction note
-
card: logo + name lockup, latest-model framing, humanized prose
-
logo: final MAREA tile (4-color ship version)
-
policy: weights note states E35 training + E36 calibration, no retraining
-
card: prose pass + latest-model framing (provenance kept, version-diff frame removed)
-
card: repo renamed to github.com/Gjusev/MK-SH-1 + link to the honest Use cases section (metrics unchanged)
-
v1.2: embed inference-policy.json (temperature + abstention + response semantics + provenance)
-
v1.2: calibration (T=7.653) + abstention policy (tau=0.79, 90% coverage, <=5% selective risk) on v1.1 weights unchanged — card updated, limitation removed
-
Author attribution: Youssef Ouhaghi Ahmian (@Gjusev )
-
Rename: mokka-1 -> MK-SH-1
18 days ago
-
Upload README.md with huggingface_hub
-
Upload model.safetensors with huggingface_hub
-
Upload tokenizer_config.json with huggingface_hub
-
Upload tokenizer.json with huggingface_hub
-
Upload README.md with huggingface_hub
-
Upload model.safetensors with huggingface_hub
-
Upload ARCHITECTURE.md with huggingface_hub
-
initial commit
