├── .gitignore
├── README.md
├── create_dataset.ipynb
├── filter_and_push.ipynb
├── image.png
├── synthetic_data.jsonl
└── synthetic_data_filtered.jsonl
/.gitignore:
--------------------------------------------------------------------------------
1 | /.env
--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 | # Creating Synthetic Dataset Using Llama 3.1 405B and Nemotron 4
2 |
3 | In this notebook we will use the following structure to create a synthetic dataset of Intructions and Git Commands.
4 |
5 |
6 | 
7 |
8 | We will create a set of instructions related to git queries in natural language, then we will generate the response for each instruction.
9 |
10 | The instruction/response pairs will be passed to a reward model, Nemotron 4, to filter out any bad pairs.
11 |
12 | Finally, the dataset will be pushed to HuggingFace.
13 |
14 | ## Dataset
15 | The final work can be viewed here:
16 | https://huggingface.co/datasets/hesamsheikh/git-prompt
--------------------------------------------------------------------------------
/filter_and_push.ipynb:
--------------------------------------------------------------------------------
1 | {
2 | "cells": [
3 | {
4 | "cell_type": "code",
5 | "execution_count": 1,
6 | "metadata": {},
7 | "outputs": [
8 | {
9 | "data": {
10 | "text/plain": [
11 | "True"
12 | ]
13 | },
14 | "execution_count": 1,
15 | "metadata": {},
16 | "output_type": "execute_result"
17 | }
18 | ],
19 | "source": [
20 | "from rich import print\n",
21 | "import os\n",
22 | "from openai import OpenAI\n",
23 | "from datasets import Dataset, DatasetDict, load_dataset\n",
24 | "import json\n",
25 | "from dotenv import load_dotenv\n",
26 | "\n",
27 | "# Load environment variables from .env file\n",
28 | "load_dotenv()"
29 | ]
30 | },
31 | {
32 | "cell_type": "code",
33 | "execution_count": 2,
34 | "metadata": {},
35 | "outputs": [],
36 | "source": [
37 | "client = OpenAI(\n",
38 | " base_url=\"https://integrate.api.nvidia.com/v1\",\n",
39 | " api_key=os.environ[\"NVIDIA_API_KEY\"]\n",
40 | ")"
41 | ]
42 | },
43 | {
44 | "cell_type": "code",
45 | "execution_count": 3,
46 | "metadata": {},
47 | "outputs": [
48 | {
49 | "data": {
50 | "text/html": [
51 | "[\n",
52 | " ChatCompletionTokenLogprob(token='helpfulness', bytes=None, logprob=4.15625, top_logprobs=[]),\n",
53 | " ChatCompletionTokenLogprob(token='correctness', bytes=None, logprob=4.125, top_logprobs=[]),\n",
54 | " ChatCompletionTokenLogprob(token='coherence', bytes=None, logprob=4.09375, top_logprobs=[]),\n",
55 | " ChatCompletionTokenLogprob(token='complexity', bytes=None, logprob=0.55078125, top_logprobs=[]),\n",
56 | " ChatCompletionTokenLogprob(token='verbosity', bytes=None, logprob=0.52734375, top_logprobs=[])\n",
57 | "]\n",
58 | "
\n"
59 | ],
60 | "text/plain": [
61 | "\u001b[1m[\u001b[0m\n",
62 | " \u001b[1;35mChatCompletionTokenLogprob\u001b[0m\u001b[1m(\u001b[0m\u001b[33mtoken\u001b[0m=\u001b[32m'helpfulness'\u001b[0m, \u001b[33mbytes\u001b[0m=\u001b[3;35mNone\u001b[0m, \u001b[33mlogprob\u001b[0m=\u001b[1;36m4\u001b[0m\u001b[1;36m.15625\u001b[0m, \u001b[33mtop_logprobs\u001b[0m=\u001b[1m[\u001b[0m\u001b[1m]\u001b[0m\u001b[1m)\u001b[0m,\n",
63 | " \u001b[1;35mChatCompletionTokenLogprob\u001b[0m\u001b[1m(\u001b[0m\u001b[33mtoken\u001b[0m=\u001b[32m'correctness'\u001b[0m, \u001b[33mbytes\u001b[0m=\u001b[3;35mNone\u001b[0m, \u001b[33mlogprob\u001b[0m=\u001b[1;36m4\u001b[0m\u001b[1;36m.125\u001b[0m, \u001b[33mtop_logprobs\u001b[0m=\u001b[1m[\u001b[0m\u001b[1m]\u001b[0m\u001b[1m)\u001b[0m,\n",
64 | " \u001b[1;35mChatCompletionTokenLogprob\u001b[0m\u001b[1m(\u001b[0m\u001b[33mtoken\u001b[0m=\u001b[32m'coherence'\u001b[0m, \u001b[33mbytes\u001b[0m=\u001b[3;35mNone\u001b[0m, \u001b[33mlogprob\u001b[0m=\u001b[1;36m4\u001b[0m\u001b[1;36m.09375\u001b[0m, \u001b[33mtop_logprobs\u001b[0m=\u001b[1m[\u001b[0m\u001b[1m]\u001b[0m\u001b[1m)\u001b[0m,\n",
65 | " \u001b[1;35mChatCompletionTokenLogprob\u001b[0m\u001b[1m(\u001b[0m\u001b[33mtoken\u001b[0m=\u001b[32m'complexity'\u001b[0m, \u001b[33mbytes\u001b[0m=\u001b[3;35mNone\u001b[0m, \u001b[33mlogprob\u001b[0m=\u001b[1;36m0\u001b[0m\u001b[1;36m.55078125\u001b[0m, \u001b[33mtop_logprobs\u001b[0m=\u001b[1m[\u001b[0m\u001b[1m]\u001b[0m\u001b[1m)\u001b[0m,\n",
66 | " \u001b[1;35mChatCompletionTokenLogprob\u001b[0m\u001b[1m(\u001b[0m\u001b[33mtoken\u001b[0m=\u001b[32m'verbosity'\u001b[0m, \u001b[33mbytes\u001b[0m=\u001b[3;35mNone\u001b[0m, \u001b[33mlogprob\u001b[0m=\u001b[1;36m0\u001b[0m\u001b[1;36m.52734375\u001b[0m, \u001b[33mtop_logprobs\u001b[0m=\u001b[1m[\u001b[0m\u001b[1m]\u001b[0m\u001b[1m)\u001b[0m\n",
67 | "\u001b[1m]\u001b[0m\n"
68 | ]
69 | },
70 | "metadata": {},
71 | "output_type": "display_data"
72 | }
73 | ],
74 | "source": [
75 | "messages = [\n",
76 | " {\n",
77 | " \"role\": \"user\",\n",
78 | " \"content\": \"What is the capital of Iran?\"\n",
79 | " },\n",
80 | " {\n",
81 | " \"role\": \"assistant\",\n",
82 | " \"content\": \"The capital of Iran is Tehran.\"\n",
83 | " },\n",
84 | "]\n",
85 | "\n",
86 | "response = client.chat.completions.create(\n",
87 | " model=\"nvidia/nemotron-4-340b-reward\",\n",
88 | " messages=messages,\n",
89 | ")\n",
90 | "\n",
91 | "print(response.choices[0].logprobs.content)"
92 | ]
93 | },
94 | {
95 | "cell_type": "code",
96 | "execution_count": 4,
97 | "metadata": {},
98 | "outputs": [
99 | {
100 | "data": {
101 | "text/html": [
102 | "{\n",
103 | " 'helpfulness': 4.15625,\n",
104 | " 'correctness': 4.125,\n",
105 | " 'coherence': 4.09375,\n",
106 | " 'complexity': 0.55078125,\n",
107 | " 'verbosity': 0.52734375\n",
108 | "}\n",
109 | "\n"
110 | ],
111 | "text/plain": [
112 | "\u001b[1m{\u001b[0m\n",
113 | " \u001b[32m'helpfulness'\u001b[0m: \u001b[1;36m4.15625\u001b[0m,\n",
114 | " \u001b[32m'correctness'\u001b[0m: \u001b[1;36m4.125\u001b[0m,\n",
115 | " \u001b[32m'coherence'\u001b[0m: \u001b[1;36m4.09375\u001b[0m,\n",
116 | " \u001b[32m'complexity'\u001b[0m: \u001b[1;36m0.55078125\u001b[0m,\n",
117 | " \u001b[32m'verbosity'\u001b[0m: \u001b[1;36m0.52734375\u001b[0m\n",
118 | "\u001b[1m}\u001b[0m\n"
119 | ]
120 | },
121 | "metadata": {},
122 | "output_type": "display_data"
123 | }
124 | ],
125 | "source": [
126 | "def get_scores_from_response(score_response_template):\n",
127 | " logprobs = score_response_template.choices[0].logprobs.content\n",
128 | " score_dict = {}\n",
129 | " for score in logprobs:\n",
130 | " score_dict[score.token] = score.logprob\n",
131 | " return score_dict\n",
132 | "\n",
133 | "print(get_scores_from_response(response))"
134 | ]
135 | },
136 | {
137 | "cell_type": "code",
138 | "execution_count": 5,
139 | "metadata": {},
140 | "outputs": [],
141 | "source": [
142 | "def get_response_and_scores(client, model, question, response_content):\n",
143 | " messages = [\n",
144 | " {\n",
145 | " \"role\": \"user\",\n",
146 | " \"content\": question\n",
147 | " },\n",
148 | " {\n",
149 | " \"role\": \"assistant\",\n",
150 | " \"content\": response_content\n",
151 | " },\n",
152 | " ]\n",
153 | "\n",
154 | " response = client.chat.completions.create(\n",
155 | " model=model,\n",
156 | " messages=messages,\n",
157 | " )\n",
158 | "\n",
159 | " scores = get_scores_from_response(response)\n",
160 | " return scores"
161 | ]
162 | },
163 | {
164 | "cell_type": "code",
165 | "execution_count": 64,
166 | "metadata": {},
167 | "outputs": [
168 | {
169 | "data": {
170 | "text/plain": [
171 | "335"
172 | ]
173 | },
174 | "execution_count": 64,
175 | "metadata": {},
176 | "output_type": "execute_result"
177 | }
178 | ],
179 | "source": [
180 | "import json\n",
181 | "\n",
182 | "def read_synthetic_data(file_path):\n",
183 | " data_list = []\n",
184 | " with open(file_path, 'r') as file:\n",
185 | " for line in file:\n",
186 | " entry = json.loads(line)\n",
187 | " one_line_dict = {}\n",
188 | " one_line_dict[\"instruction\"] = entry['instruction']\n",
189 | " one_line_dict[\"response\"] = entry['responses']\n",
190 | " data_list.append(one_line_dict)\n",
191 | " return data_list\n",
192 | "\n",
193 | "synthetic_data = read_synthetic_data(\"synthetic_data.jsonl\")\n",
194 | "len(synthetic_data)"
195 | ]
196 | },
197 | {
198 | "cell_type": "code",
199 | "execution_count": 7,
200 | "metadata": {},
201 | "outputs": [
202 | {
203 | "data": {
204 | "text/plain": [
205 | "{'instruction': 'Create a new branch for my feature',\n",
206 | " 'response': \"To create a new branch for your feature, use the following Git command:\\n\\n`git branch feature/your-feature-name`\\n\\nReplace `your-feature-name` with a descriptive name for your feature. This will create a new branch, but you'll still be on the current branch.\\n\\nTo switch to the new branch, use:\\n\\n`git checkout feature/your-feature-name`\\n\\nAlternatively, you can create and switch to the new branch in one step with:\\n\\n`git checkout -b feature/your-feature-name`\"}"
207 | ]
208 | },
209 | "execution_count": 7,
210 | "metadata": {},
211 | "output_type": "execute_result"
212 | }
213 | ],
214 | "source": [
215 | "synthetic_data[0]"
216 | ]
217 | },
218 | {
219 | "cell_type": "code",
220 | "execution_count": 8,
221 | "metadata": {},
222 | "outputs": [],
223 | "source": [
224 | "def process_instruction_response_score(client, model, synthetic_data):\n",
225 | " score_list = []\n",
226 | " for instruction_response_pair in synthetic_data:\n",
227 | " instruction = instruction_response_pair[\"instruction\"]\n",
228 | " response = instruction_response_pair[\"response\"]\n",
229 | " score = get_response_and_scores(client, model, instruction, response)\n",
230 | " score_list.append(score)\n",
231 | " return score_list"
232 | ]
233 | },
234 | {
235 | "cell_type": "code",
236 | "execution_count": 9,
237 | "metadata": {},
238 | "outputs": [],
239 | "source": [
240 | "score_list = process_instruction_response_score(client, \"nvidia/nemotron-4-340b-reward\", synthetic_data)"
241 | ]
242 | },
243 | {
244 | "cell_type": "code",
245 | "execution_count": 65,
246 | "metadata": {},
247 | "outputs": [
248 | {
249 | "data": {
250 | "text/plain": [
251 | "322"
252 | ]
253 | },
254 | "execution_count": 65,
255 | "metadata": {},
256 | "output_type": "execute_result"
257 | }
258 | ],
259 | "source": [
260 | "helpfulness_THRESHOLD = 3\n",
261 | "verbosity_THRESHOLD = 2.5\n",
262 | "synthetic_data = [data for i, data in enumerate(synthetic_data) \n",
263 | " if not (score_list[i][\"helpfulness\"] < helpfulness_THRESHOLD or \n",
264 | " score_list[i][\"verbosity\"] > verbosity_THRESHOLD)]\n",
265 | "len(synthetic_data)"
266 | ]
267 | },
268 | {
269 | "cell_type": "code",
270 | "execution_count": 67,
271 | "metadata": {},
272 | "outputs": [],
273 | "source": [
274 | "with open('synthetic_data_filtered.jsonl', 'w') as f:\n",
275 | " for item in synthetic_data:\n",
276 | " f.write(json.dumps(item))\n",
277 | " f.write('\\n')"
278 | ]
279 | },
280 | {
281 | "cell_type": "code",
282 | "execution_count": 71,
283 | "metadata": {},
284 | "outputs": [
285 | {
286 | "data": {
287 | "application/vnd.jupyter.widget-view+json": {
288 | "model_id": "56874f3fc89e443f97b34bf2e01600b8",
289 | "version_major": 2,
290 | "version_minor": 0
291 | },
292 | "text/plain": [
293 | "VBox(children=(HTML(value='