New model from the upstream. R=24 with only QV tuned.

adapter_config.json

@@ -6,17 +6,16 @@
     false,
     true
   ],
-  "fan_in_fan_out": false,
+  "fan_in_fan_out": true,
   "inference_mode": true,
   "lora_alpha": 32,
   "lora_dropout": 0.1,
-  "merge_weights": true,
+  "merge_weights": false,
   "modules_to_save": null,
   "peft_type": "LORA",
   "r": 24,
   "target_modules": [
-    "query_key_value",
-    "lm_head"
+    "query_key_value"
   ],
   "task_type": "CAUSAL_LM"
-}
+}

adapter_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:34d6a0040d59ab2308b57dc5799452ea1d57c68290186756091625d59bfcc578
+oid sha256:3f27065da4fb24beb19f928fe9699ed6c31576a2c60f30a821e6e28047fb2544
 size 44061081