Upload builder.ipynb

builder.ipynb

@@ -0,0 +1,352 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "This is Arjun's attempt at building a long-content benchmark based on HumanEvalPlus,\n",
+    "as imagined by Leandro."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 88,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import datasets\n",
+    "import random\n",
+    "import bounded_subprocess\n",
+    "import tempfile\n",
+    "from pathlib import Path"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Let's start by thanking Loubna for uploading this to the Hub."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 41,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Found cached dataset parquet (/home/arjun/.cache/huggingface/datasets/loubnabnl___parquet/loubnabnl--humaneval_plus-d3a2da5c53783cd1/0.0.0/2a3b91fbd88a2c90d1dbbb32b460cf621d31bd5b05b934492fdef7d8d6f236ec)\n"
+     ]
+    }
+   ],
+   "source": [
+    "humanevalplus = datasets.load_dataset(\"loubnabnl/humaneval_plus\", split=\"train\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "The tests in HumanEval are in the same style as HumanEval:\n",
+    "\n",
+    "```\n",
+    "def check(candidate):\n",
+    "    assert candidate(x) == y\n",
+    "    ...\n",
+    "```\n",
+    "\n",
+    "The code below extracts the assrtions, unindents them, and renamed `candidate`\n",
+    "to the name of the function being tested. Moreover, not all lines are simple\n",
+    "assertions, so we skip over them. There is a possibiliby of error: an assertion\n",
+    "may span several lines. But, it's fairly unlikely, and the models we are testing\n",
+    "shouldn't fall apart on a little noise like that.\n",
+    "\n",
+    "Finally, we strip out the docstring from the prompt.\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 71,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Loading cached processed dataset at /home/arjun/.cache/huggingface/datasets/loubnabnl___parquet/loubnabnl--humaneval_plus-d3a2da5c53783cd1/0.0.0/2a3b91fbd88a2c90d1dbbb32b460cf621d31bd5b05b934492fdef7d8d6f236ec/cache-25a2ca89ddab56cb.arrow\n"
+     ]
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "2bdeec04efc94b4bb45601a963eff47e",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Filter:   0%|          | 0/164 [00:00<?, ? examples/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "def extract_and_unindent(s, entrypoint):\n",
+    "    idx = s.find(\"def check(candidate):\")\n",
+    "    if idx == -1:\n",
+    "        return None\n",
+    "    extracted = s[idx+len(\"def check(candidate):\"):]\n",
+    "    lines = extracted.split(\"\\n\")\n",
+    "    tests = [ ]\n",
+    "    for line in lines:\n",
+    "        if line == \"\":\n",
+    "            continue\n",
+    "        if not line.startswith(\"    assert\"):\n",
+    "            continue\n",
+    "        tests.append(line.strip().replace(\"candidate(\", entrypoint + \"(\"))\n",
+    "    return tests\n",
+    "\n",
+    "def clean_item(item):\n",
+    "    tests = extract_and_unindent(item[\"test\"], item[\"entry_point\"])\n",
+    "    prompt = item[\"prompt\"][:item[\"prompt\"].find(\"\\n    \")]\n",
+    "    return {\n",
+    "        \"tests\": tests,\n",
+    "        \"prompt\": prompt,\n",
+    "        \"canonical\": item[\"canonical_solution\"].strip()\n",
+    "    }\n",
+    "\n",
+    "processed_humaneval_plus = humanevalplus.map(clean_item).filter(lambda item: len(item[\"tests\"]) > 0)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Given `processed_humaneval_plus`, we turn each item into a benchmark:\n",
+    "\n",
+    "- `prompt` has several assertions, including distractors, in random order and\n",
+    "  concludes with a function signature `def f(x):`.\n",
+    "- `size` is the length of the prompt in characters.\n",
+    "- `target_tests` are the subset of the assertions that test `f`.\n",
+    "- `canonical_prompt` is the prompt without distractors and assertions\n",
+    "- `canonical_solution` is a canonical solution that should pass the tests."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 101,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def build_benchmark(ds, other_indices, target_index):\n",
+    "    canonical_prompt = ds[target_index][\"prompt\"]\n",
+    "    canonical_solution = ds[target_index][\"canonical\"]\n",
+    "\n",
+    "    tests = []\n",
+    "    tests.extend(ds[target_index][\"tests\"])\n",
+    "    for ix in other_indices:\n",
+    "        tests.extend(ds[ix][\"tests\"])\n",
+    "    random.shuffle(tests)\n",
+    "    prompt = \"\\n\".join(tests)\n",
+    "    prompt = prompt + \"\\n\\n\" + canonical_prompt\n",
+    "    target_tests = \"\\n\".join(ds[target_index][\"tests\"])\n",
+    "    return {\n",
+    "        \"prompt\": prompt, \n",
+    "        \"target_tests\": target_tests,\n",
+    "        \"canonical_prompt\": canonical_prompt,\n",
+    "        \"canonical_solution\": \"\\n    \" + canonical_solution,\n",
+    "        \"size\": len(prompt)\n",
+    "    }\n",
+    "\n",
+    "def random_benchmark(ds, size: int):\n",
+    "    assert size > 0\n",
+    "    indices = random.sample(range(len(ds)), size)\n",
+    "    return build_benchmark(ds, indices[1:], indices[0])\n",
+    "\n",
+    "def validate_benchmark(item):\n",
+    "    program = item[\"canonical_prompt\"] + item[\"canonical_solution\"] + \"\\n\\n\" + item[\"target_tests\"]\n",
+    "    with tempfile.NamedTemporaryFile(suffix=\".py\", delete=True) as f:\n",
+    "        Path(f.name).write_text(program)\n",
+    "        r = bounded_subprocess.run([\"python3\", f.name])\n",
+    "        return r.exit_code == 0"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "This is a decent way to prompt an instruction-tuned model, but we aren't going to do it right now."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 83,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "These are several assertions:\n",
+      "\n",
+      "```\n",
+      "assert words_string(\"One,, two, three, four, five, six,\") == [\"One\", \"two\", \"three\", \"four\", \"five\", \"six\"]\n",
+      "assert words_string(\"One, two, three, four, five, six\") == [\"One\", \"two\", \"three\", \"four\", \"five\", \"six\"]\n",
+      "assert separate_paren_groups('() (()) ((())) (((())))') == [\n",
+      "assert separate_paren_groups('(()(())((())))') == [\n",
+      "assert True, \"This prints if this assert fails 2 (also good for debugging!)\"\n",
+      "assert words_string(\"\") == []\n",
+      "assert words_string(\"Hi, my name\") == [\"Hi\", \"my\", \"name\"]\n",
+      "assert words_string(\"Hi, my name is John\") == [\"Hi\", \"my\", \"name\", \"is\", \"John\"]\n",
+      "assert True, \"This prints if this assert fails 1 (good for debugging!)\"\n",
+      "assert separate_paren_groups('(()()) ((())) () ((())()())') == [\n",
+      "assert words_string(\"ahmed     , gamal\") == [\"ahmed\", \"gamal\"]\n",
+      "assert separate_paren_groups('( ) (( )) (( )( ))') == ['()', '(())', '(()())']\n",
+      "```\n",
+      "\n",
+      "Complete the following function so that the assertions pass:\n",
+      "\n",
+      "```\n",
+      "def words_string(s):\n",
+      "```\n"
+     ]
+    }
+   ],
+   "source": [
+    "b = random_benchmark(processed_humaneval_plus, 2)\n",
+    "b_assertions = b[\"prompt\"].split(\"\\n\")\n",
+    "b_signature = b_assertions[-1]\n",
+    "b_assertions = \"\\n\".join(b_assertions[:-1]).rstrip()\n",
+    "print(f\"These are several assertions:\\n\\n```\\n{b_assertions}\\n```\\n\\nComplete the following function so that the assertions pass:\\n\\n```\\n{b_signature}\\n```\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Now we build a benchmark of varying size."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 102,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Failed to generate benchmark of size 40\n"
+     ]
+    }
+   ],
+   "source": [
+    "items = [ ] \n",
+    "for size in [10, 20, 40, 80, 160]:\n",
+    "    for i in range(5):\n",
+    "        b = random_benchmark(processed_humaneval_plus, size)\n",
+    "        if validate_benchmark(b):\n",
+    "            items.append(b)\n",
+    "        else:\n",
+    "            print(f\"Failed to generate benchmark of size {size}\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 107,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "a5a96a8665534bb48490901996856341",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Pushing dataset shards to the dataset hub:   0%|          | 0/1 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "68a8489fca5a4f3e81d1a832080b157e",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Downloading metadata:   0%|          | 0.00/517 [00:00<?, ?B/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Updating downloaded metadata with the new split.\n"
+     ]
+    }
+   ],
+   "source": [
+    "longtest_benchmark = datasets.Dataset.from_list(items)\n",
+    "longtest_benchmark.push_to_hub(\"nuprl-staging/longtest_benchmark\", private=False)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "How long is the longest benchmark (in characters, not tokens):"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 108,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "64230"
+      ]
+     },
+     "execution_count": 108,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "max(longtest_benchmark[\"size\"])"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "venv",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.6"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}