add evals

evals/arc/arc_nl_stablelm-2-1_6b.json

@@ -0,0 +1,23 @@
+{
+  "results": {
+    "arc_nl": {
+      "acc": 0.28828058169375537,
+      "acc_stderr": 0.013253808186010754,
+      "acc_norm": 0.30710008554319934,
+      "acc_norm_stderr": 0.013497514452902164
+    }
+  },
+  "versions": {
+    "arc_nl": 0
+  },
+  "config": {
+    "model": "hf-auto",
+    "model_args": "pretrained=stabilityai/stablelm-2-1_6b,use_accelerate=True,device_map_option=auto,dtype=auto,load_in_8bit=True",
+    "batch_size": 4,
+    "device": "cuda",
+    "no_cache": false,
+    "limit": null,
+    "bootstrap_iters": 100000,
+    "description_dict": {}
+  }
+}

evals/hellaswag/hellaswag_nl_stablelm-2-1_6b.json

@@ -0,0 +1,23 @@
+{
+  "results": {
+    "hellaswag_nl": {
+      "acc": 0.37625472207231514,
+      "acc_stderr": 0.0050332133723218365,
+      "acc_norm": 0.4900161899622234,
+      "acc_norm_stderr": 0.005193787011918009
+    }
+  },
+  "versions": {
+    "hellaswag_nl": 1
+  },
+  "config": {
+    "model": "hf-auto",
+    "model_args": "pretrained=stabilityai/stablelm-2-1_6b,use_accelerate=True,device_map_option=auto,dtype=auto,load_in_8bit=True",
+    "batch_size": 4,
+    "device": "cuda",
+    "no_cache": false,
+    "limit": null,
+    "bootstrap_iters": 100000,
+    "description_dict": {}
+  }
+}

evals/mmlu/mmlu_nl_stablelm-2-1_6b.json

@@ -0,0 +1,23 @@
+{
+  "results": {
+    "mmlu_nl": {
+      "acc": 0.30803673066707143,
+      "acc_stderr": 0.004022082556273912,
+      "acc_norm": 0.3117553312590119,
+      "acc_norm_stderr": 0.004035399904594038
+    }
+  },
+  "versions": {
+    "mmlu_nl": 0
+  },
+  "config": {
+    "model": "hf-auto",
+    "model_args": "pretrained=stabilityai/stablelm-2-1_6b,use_accelerate=True,device_map_option=auto,dtype=auto,load_in_8bit=True",
+    "batch_size": 4,
+    "device": "cuda",
+    "no_cache": false,
+    "limit": null,
+    "bootstrap_iters": 100000,
+    "description_dict": {}
+  }
+}

evals/truthfulqa/truthfulqa_nl_stablelm-2-1_6b.json

@@ -0,0 +1,23 @@
+{
+  "results": {
+    "truthfulqa_nl": {
+      "mc1": 0.2713375796178344,
+      "mc1_stderr": 0.015880352123056246,
+      "mc2": 0.43608074786210904,
+      "mc2_stderr": 0.015073401669246872
+    }
+  },
+  "versions": {
+    "truthfulqa_nl": 1
+  },
+  "config": {
+    "model": "hf-auto",
+    "model_args": "pretrained=stabilityai/stablelm-2-1_6b,use_accelerate=True,device_map_option=auto,dtype=auto,load_in_8bit=True",
+    "batch_size": 4,
+    "device": "cuda",
+    "no_cache": false,
+    "limit": null,
+    "bootstrap_iters": 100000,
+    "description_dict": {}
+  }
+}