{ "cells": [ { "cell_type": "code", "execution_count": null, "id": "585da432", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Number of parquet files 30\n", "Reading geclm-datasets/samples/c4/20230404_102105_00007_t8w9z_3085d601-45f1-443a-b50d-8eb4812dd227\n", "Number of parquet files 30\n", "Reading geclm-datasets/samples/bigcode_python_code/20230404_102116_00007_ajvns_4e5b2899-8640-4a4c-b0cd-758662178176\n", "Number of parquet files 30\n", "Reading geclm-datasets/samples/bigcode_python_github_issues/20230404_102127_00022_yv77i_982f928f-1431-4ea7-986d-c5c5cb0f4a3f\n", "Number of parquet files 30\n", "Reading geclm-datasets/samples/bigcode_python_jupyter_markdowned_clean_dedup/20230404_102137_00026_vwcg7_3167c932-87a1-4fec-ad01-215831d0bf6e\n", "Number of parquet files 30\n", "Reading geclm-datasets/samples/books3/20230404_102143_00027_t4kwf_198fc997-b871-4e4a-b88e-3776f1cf92fe\n", "Number of parquet files 30\n", "Reading geclm-datasets/samples/gutenberg_raw/20230404_102215_00007_x3ntt_30873bfe-c94c-439a-96e2-71165570dc99\n", "Number of parquet files 30\n", "Reading geclm-datasets/samples/reddit_threaded/20230404_102241_00049_xj4uk_d7612f5a-5107-46e1-b710-47e7db95a7e6\n", "Number of parquet files 30\n", "Reading geclm-datasets/samples/enwiki_data/20230404_102246_00007_ye63c_57166ca6-f0d2-40ef-8ae7-ed4bc7ecd28d\n", "Number of parquet files 30\n", "Reading geclm-datasets/samples/s2orc_dedup/20230404_102252_00080_6ce5q_330e23f7-1270-4a52-b277-af823baf1de6\n", "Number of parquet files 30\n", "Reading geclm-datasets/samples/stackexchange2/20230404_102308_00031_qvnh6_cec28e17-f163-4a04-9fbe-dc617d9ea03e\n", "Number of parquet files 30\n", "Reading geclm-datasets/samples/commoncrawl/20230404_124237_00026_sin5w_c2e65b68-2449-47fa-be8b-a6e6e83611d0\n", "Running on local URL: http://127.0.0.1:7860\n", "\n", "To create a public link, set `share=True` in `launch()`.\n" ] }, { "data": { "text/html": [ "
" ], "text/plain": [ "" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "import math\n", "import os\n", "import random\n", "import uuid\n", "from datetime import datetime\n", "\n", "import gradio as gr\n", "import jsonlines\n", "import pyarrow as pa\n", "import s3fs\n", "from datasets import Dataset\n", "from huggingface_hub import HfApi\n", "\n", "S3 = s3fs.S3FileSystem(anon=False, key=os.getenv(\"AWS_ACCESS_KEY_ID\"), secret=os.getenv(\"AWS_SECRET_ACCESS_KEY\"))\n", "\n", "DEFAULT_SHUFFLE_BUFFER_SIZE_RATIO = 5\n", "BASE_S3_DIR = \"s3://geclm-datasets/samples/\"\n", "\n", "DATASETS = [\n", " \"c4\",\n", " \"bigcode_python_code\",\n", " \"bigcode_python_github_issues\",\n", " \"bigcode_python_jupyter_markdowned_clean_dedup\",\n", " \"books3\",\n", " \"gutenberg_raw\",\n", " \"reddit_threaded\",\n", " \"enwiki_data\",\n", " \"s2orc_dedup\",\n", " \"stackexchange2\",\n", " \"commoncrawl\",\n", "]\n", "\n", "\n", "def get_parquet_lines(dataset, sample_size=100):\n", " s3_paths = S3.glob(BASE_S3_DIR + dataset + \"/*\")\n", "\n", " if len(s3_paths) == 0:\n", " raise FileNotFoundError(f\"Nothing found at {path}\")\n", "\n", " print(\"Number of parquet files\", len(s3_paths))\n", " s3_path = random.choice(s3_paths)\n", " print(\"Reading\", s3_path)\n", " lines = []\n", "\n", " with S3.open(s3_path) as f:\n", " pf = pa.parquet.ParquetFile(f)\n", " for ix_row_group in range(pf.metadata.num_row_groups):\n", " # We load dataset by row group - 1000 rows at a time\n", " # using open_input_stream would return bytes per bytes not row per row\n", " table = pf.read_row_group(ix_row_group)\n", " lines.extend(table.to_pylist())\n", "\n", " random.shuffle(lines)\n", " return lines[:sample_size]\n", "\n", "\n", "def get_local_lines(dataset):\n", " lines = []\n", " with jsonlines.open(\"data/{}_examples_with_stats.json\".format(dataset), \"r\") as f:\n", " for line in f:\n", " lines.append(line)\n", " return lines\n", "\n", "\n", "def line_generator(lines_dict, dataset):\n", " for line in lines_dict[dataset]:\n", " yield line\n", "\n", "\n", "# Parallelize the below\n", "local_lines = {dataset: get_local_lines(dataset) for dataset in DATASETS}\n", "s3_lines = {dataset: get_parquet_lines(dataset) for dataset in DATASETS}\n", "\n", "line_generators_local = {dataset: line_generator(local_lines, dataset) for dataset in DATASETS}\n", "line_generators_s3 = {dataset: line_generator(s3_lines, dataset) for dataset in DATASETS}\n", "\n", "\n", "def send_report(sample, dataset, reason, annotator, campaign):\n", " text = sample[\"text\"]\n", " sample.pop(\"text\")\n", "\n", " sample_id = \"\"\n", " if \"id\" not in sample:\n", " if \"title\" in sample:\n", " sample_id = sample[\"title\"]\n", " else:\n", " sample_id = sample[\"id\"]\n", "\n", " with jsonlines.open(\"report.jsonl\", \"w\") as f:\n", " f.write(\n", " {\n", " \"dataset\": dataset,\n", " \"docid\": sample_id,\n", " \"text\": text,\n", " \"metadata\": sample,\n", " \"reason\": reason,\n", " \"annotator\": annotator,\n", " \"campaign\": campaign,\n", " \"timestamp\": str(datetime.now()),\n", " }\n", " )\n", "\n", " api = HfApi()\n", " api.upload_file(\n", " path_or_fileobj=\"report.jsonl\",\n", " path_in_repo=\"report-{}.jsonl\".format(uuid.uuid4()),\n", " repo_id=\"HuggingFaceGECLM/data_feedback\",\n", " repo_type=\"dataset\",\n", " token=os.environ.get(\"geclm_token\"),\n", " )\n", "\n", "\n", "description = \"\"\"\n", "GecLM annotations. All annotations are recorded in the [data_feedback](https://huggingface.co/datasets/HuggingFaceGECLM/data_feedback) dataset.\n", "\"\"\"\n", "\n", "\n", "if __name__ == \"__main__\":\n", " demo = gr.Blocks()\n", "\n", " with demo:\n", " current_sample_state = gr.State(dict())\n", "\n", " description = gr.Markdown(value=description)\n", " with gr.Row():\n", " annotator = gr.Textbox(\n", " lines=1,\n", " max_lines=1,\n", " placeholder=\"Optionally provide your name here if you'd like it to be recorded.\",\n", " label=\"Annotator\",\n", " )\n", " campaign = gr.Textbox(\n", " lines=1,\n", " max_lines=1,\n", " placeholder=\"Optionally provide the name of the annotation campagin for ease of filtering the reports.\",\n", " label=\"Annotation campaign\",\n", " )\n", " with gr.Row():\n", " dataset = gr.Dropdown(\n", " choices=DATASETS,\n", " value=\"Pick a dataset below\",\n", " label=\"Dataset\",\n", " )\n", " with gr.Row():\n", " reason_txt = gr.Textbox(\n", " label=\"Flagging reason\",\n", " placeholder=\"Provide the reason for flagging if you think the sample is bad.\",\n", " visible=False,\n", " )\n", " with gr.Row():\n", " bad_btn = gr.Button(\"Bad ❌\", visible=False)\n", " good_btn = gr.Button(\"Next ✅\", visible=False)\n", " with gr.Row():\n", " text = gr.Textbox(visible=False, label=\"Datapoint\", lines=500)\n", "\n", " def next_line(dataset):\n", " next_line = next(line_generators_s3[dataset])\n", "\n", " text_col = \"text\"\n", " if text_col not in next_line:\n", " text_col = \"content\"\n", " return [\n", " gr.update(value=next_line[text_col], visible=True),\n", " next_line,\n", " gr.update(visible=True),\n", " gr.update(visible=True),\n", " gr.update(visible=True),\n", " ]\n", "\n", " def bad_line(current_sample, dataset, reason, annotator, campaign):\n", " send_report(current_sample, dataset, reason, annotator, campaign)\n", " next_line = next(line_generators_s3[dataset])\n", " text_col = \"text\"\n", " if text_col not in next_line:\n", " text_col = \"content\"\n", " return [\n", " next_line[text_col],\n", " gr.update(\n", " value=\"\",\n", " placeholder=\"Provide the reason for flagging if you think the sample is bad.\",\n", " ),\n", " next_line,\n", " ]\n", "\n", " good_btn.click(\n", " next_line,\n", " inputs=dataset,\n", " outputs=[text, current_sample_state, reason_txt, good_btn, bad_btn],\n", " )\n", " dataset.change(\n", " next_line,\n", " inputs=dataset,\n", " outputs=[text, current_sample_state, reason_txt, good_btn, bad_btn],\n", " )\n", " bad_btn.click(\n", " bad_line,\n", " inputs=[current_sample_state, dataset, reason_txt, annotator, campaign],\n", " outputs=[text, reason_txt, current_sample_state],\n", " )\n", "\n", " demo.launch(enable_queue=False, debug=True)" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3 (ipykernel)", "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": "3.10.9" } }, "nbformat": 4, "nbformat_minor": 5 }