add evals

evals/arc/arc_nl_Mistral-7B-v0.1.json

@@ -0,0 +1,23 @@
+{
+  "results": {
+    "arc_nl": {
+      "acc": 0.41916167664670656,
+      "acc_stderr": 0.01443766613132885,
+      "acc_norm": 0.42172797262617623,
+      "acc_norm_stderr": 0.014449768025902365
+    }
+  },
+  "versions": {
+    "arc_nl": 0
+  },
+  "config": {
+    "model": "hf-auto",
+    "model_args": "pretrained=mistralai/Mistral-7B-v0.1,use_accelerate=True,device_map_option=auto,dtype=bfloat16,load_in_8bit=True",
+    "batch_size": "auto",
+    "device": "cuda",
+    "no_cache": false,
+    "limit": null,
+    "bootstrap_iters": 100000,
+    "description_dict": {}
+  }
+}

evals/arc/arc_nl_neural-chat-7b-v3-1.json

@@ -0,0 +1,23 @@
+{
+  "results": {
+    "arc_nl": {
+      "acc": 0.4576561163387511,
+      "acc_stderr": 0.014577585167374469,
+      "acc_norm": 0.43370402053036783,
+      "acc_norm_stderr": 0.014500969591642415
+    }
+  },
+  "versions": {
+    "arc_nl": 0
+  },
+  "config": {
+    "model": "hf-auto",
+    "model_args": "pretrained=Intel/neural-chat-7b-v3-1,use_accelerate=True,device_map_option=auto,dtype=bfloat16,load_in_8bit=True",
+    "batch_size": "auto",
+    "device": "cuda",
+    "no_cache": false,
+    "limit": null,
+    "bootstrap_iters": 100000,
+    "description_dict": {}
+  }
+}

evals/dutch_models.json

@@ -0,0 +1,6 @@
+{
+  "gpt-neo-1.3B-dutch": "PT",
+  "gpt-neo-125M-dutch": "PT",
+  "gpt2-large-dutch": "PT",
+  "gpt2-medium-dutch": "PT"
+}