instructions, more

README.md

@@ -95,7 +95,28 @@ configs:
 
 Sourced from https://github.com/mosaicml/llm-foundry/blob/main/scripts/eval/local_data/world_knowledge/jeopardy_all.jsonl
 
-Description: Jeopardy consists of 2,117 Jeopardy questions separated into 5 categories: 
-Literature, American History, World History, Word Origins, and Science. The model is expected 
-to give the exact correct response to the question. It was custom curated by MosaicML from a 
+Description: Jeopardy consists of 2,117 Jeopardy questions separated into 5 categories:
+Literature, American History, World History, Word Origins, and Science. The model is expected
+to give the exact correct response to the question. It was custom curated by MosaicML from a
 larger Jeopardy set available on [Huggingface](https://huggingface.co/datasets/jeopardy).
+
+## How to use
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset("soldni/jeopardy", "mosaicml_gauntlet")
+model = ...
+tokenizer = ...
+
+# Given context, try to predict the continuation
+for row in dataset:
+    input_ids = tokenizer(row['context'], return_tensors='pt').to(model.device)
+    outputs = model.generate(input_ids, max_new_tokens=100)
+    decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    correct = row['continuation'] in decoded
+    print("Gold:", row['continuation'])
+    print("Pred:", decoded)
+    print("Correct?", correct)
+    print("----")
+```

match.ipynb

@@ -258,14 +258,16 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 5,
+   "execution_count": 17,
    "metadata": {},
    "outputs": [
     {
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Missing: 0\n"
+      "Missing: 0\n",
+      "Old questions: 2220\n",
+      "New questions: 2113\n"
      ]
     }
    ],
@@ -275,6 +277,8 @@
     "\n",
     "missing =  [{**v, 'key': k} for k, v in new_questions.items() if k not in old_questions_subset]\n",
     "print(f\"Missing: {len(missing)}\")\n",
+    "print(f\"Old questions: {len(old_questions_subset)}\")\n",
+    "print(f\"New questions: {len(new_questions)}\")\n",
     "for m in missing:\n",
     "    print(m['context'])\n",
     "    print(m['continuation'])\n",
@@ -291,13 +295,21 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 10,
+   "execution_count": 19,
    "metadata": {},
    "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Mosaicml Gauntlet: 2116\n",
+      "All questions: 216930\n"
+     ]
+    },
     {
      "data": {
       "application/vnd.jupyter.widget-view+json": {
-       "model_id": "33f3bd1d30d54874b990c17e34b44bb8",
+       "model_id": "b0db52ff66c84083984ce81539bb3129",
        "version_major": 2,
        "version_minor": 0
       },
@@ -311,28 +323,21 @@
     {
      "data": {
       "application/vnd.jupyter.widget-view+json": {
-       "model_id": "17dae484d5ef4a7bb1b2204da34da4ca",
+       "model_id": "e07846cf593e47e78cc7c62d179c5baa",
        "version_major": 2,
        "version_minor": 0
       },
       "text/plain": [
-       "Creating parquet from Arrow format: 0ba [00:00, ?ba/s]"
+       "Creating parquet from Arrow format:   0%|          | 0/3 [00:00<?, ?ba/s]"
       ]
      },
      "metadata": {},
      "output_type": "display_data"
     },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "No files have been modified since last commit. Skipping to prevent empty commit.\n"
-     ]
-    },
     {
      "data": {
       "application/vnd.jupyter.widget-view+json": {
-       "model_id": "491d1a420a6246fb9d74201437e0d490",
+       "model_id": "af4036c1ed13441c9a6cc1de5558dfa3",
        "version_major": 2,
        "version_minor": 0
       },
@@ -346,7 +351,7 @@
     {
      "data": {
       "application/vnd.jupyter.widget-view+json": {
-       "model_id": "ea3ec39f1ba145c6a5457bc34912cf72",
+       "model_id": "e23eca2e0aac433baba52069b40c2e3a",
        "version_major": 2,
        "version_minor": 0
       },
@@ -357,13 +362,27 @@
      "metadata": {},
      "output_type": "display_data"
     },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "7a6982d43029428c98258352ec90bf28",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "README.md:   0%|          | 0.00/2.34k [00:00<?, ?B/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
     {
      "data": {
       "text/plain": [
-       "CommitInfo(commit_url='https://huggingface.co/datasets/soldni/jeopardy/commit/1618cf7f59f5e0a5132eb99cd43b3180e359fe78', commit_message='Upload dataset', commit_description='', oid='1618cf7f59f5e0a5132eb99cd43b3180e359fe78', pr_url=None, pr_revision=None, pr_num=None)"
+       "CommitInfo(commit_url='https://huggingface.co/datasets/soldni/jeopardy/commit/95da23525ad41487becfaf821be025ccefda6a34', commit_message='Upload dataset', commit_description='', oid='95da23525ad41487becfaf821be025ccefda6a34', pr_url=None, pr_revision=None, pr_num=None)"
       ]
      },
-     "execution_count": 10,
+     "execution_count": 19,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -372,12 +391,15 @@
     "all_questions = []\n",
     "mosaicml_gauntlet = []\n",
     "for row in old_questions:\n",
-    "    key = ' '.join([ch for ch in f\"{row['context']} {row['continuation']}\" if ch not in punctuation and ch not in whitespace])\n",
+    "    key = ''.join([ch for ch in f\"{row['context']} {row['continuation']}\" if ch not in punctuation and ch not in whitespace])\n",
     "    all_questions.append(row)\n",
     "    row['mosaicml_gauntlet'] = key in new_questions\n",
     "    if row['mosaicml_gauntlet']:\n",
     "        mosaicml_gauntlet.append(row)\n",
     "\n",
+    "print(f\"Mosaicml Gauntlet: {len(mosaicml_gauntlet)}\")\n",
+    "print(f\"All questions: {len(all_questions)}\")\n",
+    "\n",
     "mosaicml_gauntlet_dataset = datasets.Dataset.from_list(mosaicml_gauntlet)\n",
     "all_questions_dataset = datasets.Dataset.from_list(all_questions)\n",
     "\n",