OsirisPtah-Coder-v5 v4: proven mlabonne datasets + layer blacklist

README.md

@@ -5,4 +5,5 @@ tags: [abliteration, uncensored, qwen2.5, coder, osiris]
 pipeline_tag: text-generation
 ---
 # OsirisPtah-Coder-v5
-Abliterated `Qwen/Qwen2.5-Coder-7B-Instruct` — mean-difference method, 1.5x strength, 4 passes.
+Abliterated `Qwen/Qwen2.5-Coder-7B-Instruct` using proven mlabonne datasets (256+256), mean-diff method,
+1.5x strength, 4 passes, layer blacklist [0, 1, 26, 27].

abliteration_metadata.json

@@ -1,10 +1,16 @@
 {
   "base_model": "Qwen/Qwen2.5-Coder-7B-Instruct",
-  "technique": "mean_difference_weight_orthogonalization_v3",
-  "best_layer": 26,
+  "technique": "mean_diff_abliteration_v4_proven_datasets",
+  "best_layer": 20,
   "strength": 1.5,
   "num_passes": 4,
-  "num_harmful": 80,
-  "num_harmless": 81,
-  "created_at": "2026-03-03T02:20:40.426872"
+  "blacklisted_layers": [
+    0,
+    1,
+    26,
+    27
+  ],
+  "harmful_dataset": "mlabonne/harmful_behaviors",
+  "harmless_dataset": "mlabonne/harmless_alpaca",
+  "created_at": "2026-03-03T05:20:19.494796"
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:32cff08484f391e60586b71d183cf14f4932b2c23fe35511400891beda689631
+oid sha256:d09f5c17b78c793a3712f99d3ff2601850e0759b9af9848b634d0792c59dcb94
 size 15231272152

refusal_direction.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:df8111d5555de5c86d5bbb135c43ca1bab32e95b03c1eec92d4438beaf271d26
+oid sha256:cf6bfcdfe9e98d8ab9cbd00159bd3b79038904137b86e9847b0c178451b4b97f
 size 15983