v1.6

README.md

@@ -35,8 +35,8 @@ Dolma is a dataset of 3 trillion tokens from a diverse mix of web content, acade
 
 More information:
 
-- Read Dolma **announcement blogpost** [on Medium](https://soldni.medium.com/dolma-3-trillion-tokens-open-llm-corpus-9a0ff4b8da64);
-- Learn more about Dolma on its [**Data Sheet**](https://drive.google.com/file/d/12gOf5I5RytsD159nSP7iim_5zN31FCXq/view?usp=drive_link);
+<https://github.com/allenai/dolma/blob/main/docs/assets/dolma-datasheet-v0.1.pdf>
+- Read Dolma **manuscript** and its **Data Sheet** [on ArXiv](https://github.com/allenai/dolma/blob/soldni/paper/docs/assets/dolma-v1_6-20240131.pdf);
 - Review Dolma's [**ImpACT license** for medium risk artifacts](https://allenai.org/licenses/impact-mr);
 - Explore the [**open source tools**](https://github.com/allenai/dolma) we created to curate Dolma.
 - Want to request removal of personal data? Use [this form](https://forms.gle/q4BNUUxUxKwKkfdT6) to notify us of documents containing PII about a specific user.
@@ -44,103 +44,80 @@ More information:
 
 To learn more about the toolkit used to create Dolma, including how to replicate this dataset, head over our [GitHub project page](https://github.com/allenai/dolma/tree/main/docs)!
 
-## Summary Statistics
 
+## Versions
 
-|**Source**|**Type**|**Gzip files (GB)**|**Documents (millions)**|**[GPT-NeoX](https://huggingface.co/EleutherAI/gpt-neox-20b) Tokens (billions)**|
-|:---|:---:|:---:|:---:|:----:|
-|[CommonCrawl](https://commoncrawl.org/)|web|4,197|4,600|2,415|
-|[C4](https://huggingface.co/datasets/allenai/c4)|web|302|364|175|
-|[peS2o](https://huggingface.co/datasets/allenai/peS2o)|academic|150|38.8|57|
-|[The Stack](https://huggingface.co/datasets/bigcode/the-stack)|code|319|236|430|
-|[Project Gutenberg](https://www.gutenberg.org/)|books|6.6|0.052|4.8|
-|[Wikipedia](https://dumps.wikimedia.org/)|encyclopedic|5.8|6.1|3.6|
-||**Total** |**4980.4**|**5,245**|**3,084**|
+At the moment, there are five versions of Dolma available:
 
+| **Version** | **Default?** | **Release Date** | **Size** (gzip) | **Description** |
+|--|:--:|--|--|--|
+| `v1_6` | ✅ | 2024-01-31 | 5.4 TB | The latest version of Dolma, with 3 trillion tokens from a diverse mix of web content, academic publications, code, books, and encyclopedic materials. |
+| `v1_6-sample` | | 2024-01-31 | 16.4 GB | A smaller sample of Dolma, with roughly 10 billion tokens. Useful for data exploration. |
+| `v1_5` | | 2023-10-31 | 6.4 TB | The version of Dolma used to train [OLMo-1B](https://huggingface.co/allenai/OLMo-1B). Roughly 3 trillion tokens. |
+| `v1_5-sample` | | 2023-10-31 | 2.9 TB | A sample of roughly 1.9 trillion tokens used to train [OLMo-7B](https://huggingface.co/allenai/OLMo-7B) |
+| `v1` | | 2023-08-18 | 6.0 TB | The first version of Dolma. |
 
-## Download
+(Size difference between `v1_6` and previous version is due to different set of metadata included in files: we removed redundant metadata in `v1_6`.)
 
+## Summary Statistics (v1.6)
 
-The fastest way to download Dolma is to directly download the individual files across multiple threads.
-This can be achieved using wget or [aria2](https://github.com/aria2/aria2) Linux/Mac/Windows package (`sudo apt-get install aria2` on Ubuntu).
 
-For downloading individual files, simply use `wget` as follows:
+| **Source** | **Doc Type** | **UTF-8 bytes** (GB) | **Documents** (millions) | **Unicode words** (billions) | **Llama tokens** (billions) |
+|--|--|--|--|--|--|
+| Common Crawl | web pages | 9,022 | 3,370 | 1,775 | 2,281 |
+| The Stack | code| 1,043| 210 | 260| 411 |
+| C4 | web pages | 790 | 364 | 153| 198 |
+| Reddit| social media| 339 | 377| 72| 89 |
+| PeS2o | STEM papers| 268 | 38.8| 50| 70 |
+| Project Gutenberg | books | 20.4 | 0.056 | 4.0 | 6.0 |
+| Wikipedia, Wikibooks | encyclopedic | 16.2 | 6.2 | 3.7 | 4.3 |
+| **Total** | | **11,519** | **4,367** | **2,318** | **3,059** |
 
-`wget --header 'Authorization: Bearer YOUR_HF_HUB_ACCESS_TOKEN' https://huggingface.co/datasets/allenai/dolma/resolve/main/data/peS2o/s2_v3-0000.json.gz`
 
-For downloading many files across multiple threads, first prepare a `.txt` file with the urls you would like such as via the script below:
 
-```python
-OUT_DIRECTORY = "/scratch/dolma/data"
-
-# URLs for cc_en_head
-cc_en_head_base_url = "https://huggingface.co/datasets/allenai/dolma/resolve/main/data/common-crawl/cc_en_head/cc_en_head-"
-cc_en_head_url_list = [f"{cc_en_head_base_url}{str(i).zfill(4)}.json.gz\n  dir={OUT_DIRECTORY}/cc_en_head\n  out=cc_en_head-{str(i).zfill(4)}.json.gz" for i in range(612)]
-
-# URLs for cc_en_middle
-cc_en_middle_base_url = "https://huggingface.co/datasets/allenai/dolma/resolve/main/data/common-crawl/cc_en_middle/cc_en_middle-"
-cc_en_middle_url_list = [f"{cc_en_middle_base_url}{str(i).zfill(4)}.json.gz\n  dir={OUT_DIRECTORY}/cc_en_middle\n  out=cc_en_middle-{str(i).zfill(4)}.json.gz" for i in range(777)]
-
-# URLs for cc_en_tail
-cc_en_tail_base_url = "https://huggingface.co/datasets/allenai/dolma/resolve/main/data/common-crawl/cc_en_tail/cc_en_tail-"
-cc_en_tail_url_list = [f"{cc_en_tail_base_url}{str(i).zfill(4)}.json.gz\n  dir={OUT_DIRECTORY}/cc_en_tail\n  out=cc_en_tail-{str(i).zfill(4)}.json.gz" for i in range(1493)]
-
-# URLs for s2_v3
-s2_v3_base_url = "https://huggingface.co/datasets/allenai/dolma/resolve/main/data/peS2o/s2_v3-"
-s2_v3_url_list = [f"{s2_v3_base_url}{str(i).zfill(4)}.json.gz\n  dir={OUT_DIRECTORY}/peS2o\n  out=s2_v3-{str(i).zfill(4)}.json.gz" for i in range(42)]
-
-# URLs for The Stack
-LANG_TO_FILES = {'lasso': 1, 'nsis': 1, 'literate-agda': 1, 'metal': 1, 'xojo': 1, 'max': 8, 'jupyter-notebook': 101, 'asp': 7, 'elixir': 14, 'html+erb': 19, 'julia': 22, 'dart': 63, 'ragel-in-ruby-host': 1, 'api-blueprint': 1, 'gams': 1, 'tex': 71, 'xml': 101, 'smalltalk': 17, 'cmake': 11, 'piglatin': 1, "cap'n-proto": 1, 'common-lisp': 21, 'stylus': 3, 'typescript': 101, 'jflex': 1, 'factor': 1, 'arc': 1, 'parrot-internal-representation': 1, 'aspectj': 1, 'go': 101, 'urweb': 1, 'dns-zone': 1, 'purebasic': 1, 'toml': 15, 'erlang': 11, 'hy': 1, 'component-pascal': 2, 'oz': 1, 'opa': 1, 'handlebars': 10, 'gas': 15, 'less': 17, 'gnuplot': 15, 'harbour': 1, 'vhdl': 16, 'octave': 1, 'powershell': 21, 'clips': 1, 'fish': 1, 'prolog': 1, 'sparql': 1, 'objective-j': 1, 'scaml': 1, 'twig': 20, 'gettext-catalog': 101, 'purescript': 2, 'vala': 1, 'gosu': 1, 'apacheconf': 1, 'xc': 1, 'lean': 3, 'mako': 1, 'r': 4, 'unrealscript': 1, 'solidity': 21, 'pike': 1, 'cartocss': 1, 'maple': 1, 'graphql': 3, 'unity3d-asset': 101, 'swift': 101, 'dockerfile': 13, 'digital-command-language': 1, 'scala': 83, 'sqf': 2, 'logtalk': 1, 'coq': 1, 'shellsession': 1, 'befunge': 1, 'nu': 1, 'ecere-projects': 1, 'zimpl': 1, 'shen': 1, 'golo': 1, 'web-ontology-language': 12, 'sas': 2, 'uno': 1, 'livescript': 1, 'literate-haskell': 1, 'clojure': 8, 'perl6': 1, 'zig': 3, 'liquid': 2, 'ec': 1, 'blitzbasic': 1, 'sql': 101, 'http': 2, 'xproc': 1, 'kit': 1, 'textile': 1, 'netlinx': 1, 'propeller-spin': 1, 'cython': 5, 'realbasic': 1, 'dogescript': 1, 'llvm': 9, 'pawn': 1, 'groff': 40, 'html+django': 3, 'csound': 1, 'd': 1, 'agda': 2, 'css': 101, 'yacc': 7, 'robotframework': 1, 'kotlin': 101, 'grace': 1, 'abap': 2, 'blitzmax': 1, 'webassembly': 3, 'ampl': 1, 'postscript': 16, 'nit': 1, 'gentoo-eclass': 1, 'xpages': 1, 'linker-script': 2, 'yang': 3, 'jade': 4, 'standard-ml': 6, 'javascript': 101, 'moonscript': 1, 'mtml': 1, 'saltstack': 1, 'freemarker': 5, 'ston': 1, 'html+eex': 1, 'xs': 1, 'c++': 101, 'matlab': 1, 'm4': 2, 'xbase': 1, 'perl': 37, 'emacs-lisp': 7, 'bison': 1, 'slim': 2, 'grammatical-framework': 1, 'rdoc': 1, 'nix': 10, 'clean': 1, 'module-management-system': 1, 'nimrod': 6, 'raml': 1, 'forth': 1, 'squirrel': 1, 'alloy': 1, 'opencl': 3, 'c': 101, 'sass': 4, 'eiffel': 2, 'papyrus': 1, 'html': 109, 'java': 101, 'hcl': 14, 'isabelle': 2, 'markdown': 101, 'gentoo-ebuild': 2, 'objdump': 1, 'emberscript': 1, 'text': 101, 'bro': 1, 'opal': 1, 'haskell': 35, 'mupad': 1, 'desktop': 1, 'modelica': 2, 'coldfusion-cfc': 2, 'fantom': 1, 'glsl': 10, 'ocaml': 16, 'nesc': 2, 'scheme': 7, 'crystal': 5, 'tcsh': 1, 'c2hs-haskell': 1, 'idris': 1, 'logos': 4, 'coffeescript': 13, 'g-code': 10, 'sage': 1, 'haml': 4, 'tcl': 7, 'smt': 5, 'ox': 1, 'chuck': 1, 'xquery': 1, 'batchfile': 7, 'pod': 2, 'xtend': 1, 'restructuredtext': 61, 'rmarkdown': 1, 'turtle': 33, 'jsx': 45, 'protocol-buffer': 8, "ren'py": 2, 'diff': 32, 'slash': 1, 'darcs-patch': 1, 'numpy': 1, 'augeas': 1, 'wisp': 1, 'edn': 15, 'ooc': 1, 'bitbake': 2, 'labview': 1, 'inform-7': 1, 'rust': 101, 'creole': 1, 'apl': 1, 'arduino': 11, 'openscad': 2, 'cuda': 9, 'thrift': 1, 'yaml': 101, 'fancy': 1, 'coldfusion': 1, 'python': 101, 'clarion': 1, 'glyph': 1, 'parrot': 1, 'lookml': 1, 'java-server-pages': 19, 'oxygene': 1, 'flux': 1, 'scilab': 1, 'groovy-server-pages': 2, 'rhtml': 1, 'eagle': 52, 'parrot-assembly': 1, 'igor-pro': 1, 'webidl': 1, 'bluespec': 1, 'unified-parallel-c': 1, 'smali': 38, 'haxe': 9, 'ada': 7, 'lua': 48, 'pascal': 21, 'html+php': 6, 'irc-log': 1, 'x10': 1, 'netlogo': 1, 'ioke': 1, 'dm': 1, 'self': 1, 'elm': 5, 'ats': 1, 'brainfuck': 1, 'mask': 1, 'rouge': 1, 'turing': 1, 'lex': 2, 'gap': 1, 'pogoscript': 1, 'kicad': 30, 'io': 1, 'objective-c++': 8, 'qml': 4, 'redcode': 1, 'autoit': 2, 'processing': 4, 'systemverilog': 6, 'gdscript': 5, 'f-sharp': 12, 'fortran': 23, 'monkey': 1, 'c-sharp': 101, 'xslt': 9, 'viml': 6, 'renderscript': 1, 'scss': 84, 'cucumber': 4, 'verilog': 1, 'genshi': 1, 'racket': 1, 'krl': 1, 'actionscript': 10, 'pan': 1, 'cirru': 1, 'chapel': 1, 'pure-data': 2, 'm': 1, 'applescript': 1, 'inno-setup': 1, 'volt': 1, 'myghty': 1, 'groovy': 17, 'ags-script': 1, 'mirah': 1, 'lsl': 1, 'brightscript': 1, 'python-traceback': 1, 'sourcepawn': 2, 'maxscript': 1, 'zephir': 1, 'supercollider': 1, 'mathematica': 20, 'awk': 1, 'autohotkey': 2, 'lfe': 1, 'ruby': 101, 'visual-basic': 20, 'ini': 59, 'red': 1, 'omgrofl': 1, 'idl': 1, 'rebol': 1, 'vue': 101, 'ninja': 2, 'ecl': 1, 'lolcode': 1, 'tea': 1, 'txl': 1, 'smarty': 9, 'vcl': 1, 'php': 101, 'literate-coffeescript': 1, 'click': 1, 'pony': 1, 'mediawiki': 5, 'stata': 5, 'stan': 1, 'nginx': 1, 'asciidoc': 16, 'antlr': 1, 'cobol': 1, 'org': 5, 'latte': 1, 'makefile': 32, 'ceylon': 1, 'graphviz-(dot)': 13, 'lilypond': 1, 'dylan': 1, 'qmake': 1, 'muf': 1, 'j': 1, 'pov-ray-sdl': 1, 'jasmin': 1, 'shell': 73, 'cycript': 1, 'boo': 1, 'hlsl': 2}
-stack_base_url = "https://huggingface.co/datasets/allenai/dolma/resolve/main/data/stack-code/"
-stack_url_list = []
-for lang, num_files in sorted(LANG_TO_FILES.items()):
-    for i in range(num_files):
-        stack_url_list.append(f"{stack_base_url}{lang}/v3-{str(i).zfill(4)}.json.gz\n  dir={OUT_DIRECTORY}/stack-code/{lang}\n  out=v3-{str(i).zfill(4)}.json.gz")
-
-# Combine all URL lists
-all_url_list = cc_en_head_url_list + cc_en_middle_url_list + cc_en_tail_url_list + s2_v3_url_list + stack_url_list
-
-out = open("files.txt", "a")
-# Print the combined list of URLs
-for i, url in enumerate(all_url_list):
-    out.write(url + "\n")
-```
+## Download
 
-Then you can download them all in parallel using:
+The fastest way to download Dolma is to clone this repository and use the files in the `url` directory.
+We recommend using wget in parallel mode to download the files. For example:
 
-`aria2c --input-file files.txt --header 'Authorization: Bearer YOUR_HF_HUB_ACCESS_TOKEN'`
+```bash
+DATA_DIR="<path_to_your_data_directory>"
+PARALLEL_DOWNLOADS="<number_of_parallel_downloads>"
+DOLMA_VERSION="<version_of_dolma_to_download>"
 
-You can also add `-s` to increase the number of connections, e.g. `-s 10` (defaults to 5).
+git clone https://huggingface.co/datasets/allenai/dolma
+mkdir -p "${DATA_DIR}"
 
 
+cat "dolma/urls/${DOLMA_VERSION}.txt" | xargs -n 1 -P "${PARALLEL_DOWNLOADS}" wget -q -P "$DATA_DIR"
+```
 
-To get the exact file counts that are used for The Stack in the above script (`LANG_TO_FILES`), you can follow the below:
+Then, to load this data using HuggingFace's `datasets` library, you can use the following code:
 
-Fetch all files (does not download them, so should be fast): `GIT_LFS_SKIP_SMUDGE=1 git clone git@hf.co:datasets/allenai/dolma.git`
-Then run:
 ```python
 import os
+from datasets import load_dataset
 
-directory = "dolma/data/stack-code"
-folder_dict = {}
-
-for folder in os.listdir(directory):
-    folder_path = os.path.join(directory, folder)
-    if os.path.isdir(folder_path):
-        file_count = len([f for f in os.listdir(folder_path) if os.path.isfile(os.path.join(folder_path, f))])
-        folder_dict[folder] = file_count
-
-print(folder_dict)
+os.environ["DATA_DIR"] = "<path_to_your_data_directory>"
+dataset = load_dataset("allenai/dolma", split="train")
 ```
 
 ## Bibtex
 
 If you use our dataset or tooling, please cite us at:
 
-```
+```bibtex
 @article{dolma,
   title = {{Dolma: An Open Corpus of Three Trillion Tokens for Language Model Pretraining Research}},
-  author = {Luca Soldaini and Rodney Kinney and Akshita Bhagia and Dustin Schwenk and David Atkinson and Russell Authur and Ben Bogin and Khyathi Chandu and Jennifer Dumas and Yanai Elazar and Valentin Hofmann and Ananya Harsh Jha and Sachin Kumar and Li Lucy and Xinxi Lyu and Ian Magnusson and Jacob Morrison and Niklas Muennighoff and Aakanksha Naik and Crystal Nam and Matthew E. Peters and Abhilasha Ravichander and Kyle Richardson and Zejiang Shen and Emma Strubell and Nishant Subramani and Oyvind Tafjord and Evan Pete Walsh and Hannaneh Hajishirzi and Noah A. Smith and Luke Zettlemoyer and Iz Beltagy and Dirk Groeneveld and Jesse Dodge and Kyle Lo},
+  author = {
+    Luca Soldaini and Rodney Kinney and Akshita Bhagia and Dustin Schwenk and David Atkinson and
+    Russell Authur and Ben Bogin and Khyathi Chandu and Jennifer Dumas and Yanai Elazar and
+    Valentin Hofmann and Ananya Harsh Jha and Sachin Kumar and Li Lucy and Xinxi Lyu and Ian Magnusson and
+    Jacob Morrison and Niklas Muennighoff and Aakanksha Naik and Crystal Nam and Matthew E. Peters and
+    Abhilasha Ravichander and Kyle Richardson and Zejiang Shen and Emma Strubell and Nishant Subramani and
+    Oyvind Tafjord and Evan Pete Walsh and Hannaneh Hajishirzi and Noah A. Smith and Luke Zettlemoyer and
+    Iz Beltagy and Dirk Groeneveld and Jesse Dodge and Kyle Lo
+},
   year = {2024},
   journal={arXiv preprint},
 }

dolma.py

@@ -15,11 +15,12 @@
 # Lint as: python3
 """Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research"""
 
-
-from pathlib import Path
+import gzip
+import json
+import os
+from typing import List
 
 import datasets
-import os
 
 logger = datasets.logging.get_logger(__name__)
 
@@ -30,21 +31,24 @@ Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Re
 
 _URL_LISTS = {
     "v1": "urls/v1.txt",
-    "v1_5r1": "urls/v1_5r1.txt",
-    "v1_5r1-sample": "urls/v1_5r1-sample.txt",
-    "v1_5r2": "urls/v1_5r2.txt",
+    "v1_5": "urls/v1_5.txt",
+    "v1_5-sample": "urls/v1_5-sample.txt",
+    "v1_6": "urls/v1_6.txt",
+    "v1_6-sample": "urls/v1_6-sample.txt",
 }
 _VERSIONS = {
     "v1": "1.0.0",
-    "v1_5r1": "1.5.0",
-    "v1_5r1-sample": "1.5.0",
-    "v1_5r2": "1.5.0",
+    "v1_5": "1.5.0",
+    "v1_5-sample": "1.5.0",
+    "v1_6": "1.6.0",
+    "v1_6-sample": "1.6.0",
 }
 _DATES = {
     "v1": "(Aug 2023)",
-    "v1_5r1": "(Oct 2023)",
-    "v1_5r1-sample": "(Oct 2023)",
-    "v1_5r2": "Dolma v1.5r2 (Dec 2023)",
+    "v1_5": "(Oct 2023)",
+    "v1_5-sample": "(Oct 2023)",
+    "v1_6": "(Jan 2024)",
+    "v1_6-sample": "(Jan 2024)",
 }
 _BASE_URL = "https://olmo-data.org"
 
@@ -54,14 +58,14 @@ _CITATION = """\
 @article{dolma,
   title = {{Dolma: An Open Corpus of Three Trillion Tokens for Language Model Pretraining Research}},
   author = {
-        Luca Soldaini and Rodney Kinney and Akshita Bhagia and Dustin Schwenk and David Atkinson and
-        Russell Authur and Ben Bogin and Khyathi Chandu and Jennifer Dumas and Yanai Elazar and
-        Valentin Hofmann and Ananya Harsh Jha and Sachin Kumar and Li Lucy and Xinxi Lyu and Ian Magnusson and
-        Jacob Morrison and Niklas Muennighoff and Aakanksha Naik and Crystal Nam and Matthew E. Peters and
-        Abhilasha Ravichander and Kyle Richardson and Zejiang Shen and Emma Strubell and Nishant Subramani and
-        Oyvind Tafjord and Evan Pete Walsh and Hannaneh Hajishirzi and Noah A. Smith and Luke Zettlemoyer and
-        Iz Beltagy and Dirk Groeneveld and Jesse Dodge and Kyle Lo
-    },
+    Luca Soldaini and Rodney Kinney and Akshita Bhagia and Dustin Schwenk and David Atkinson and
+    Russell Authur and Ben Bogin and Khyathi Chandu and Jennifer Dumas and Yanai Elazar and
+    Valentin Hofmann and Ananya Harsh Jha and Sachin Kumar and Li Lucy and Xinxi Lyu and Ian Magnusson and
+    Jacob Morrison and Niklas Muennighoff and Aakanksha Naik and Crystal Nam and Matthew E. Peters and
+    Abhilasha Ravichander and Kyle Richardson and Zejiang Shen and Emma Strubell and Nishant Subramani and
+    Oyvind Tafjord and Evan Pete Walsh and Hannaneh Hajishirzi and Noah A. Smith and Luke Zettlemoyer and
+    Iz Beltagy and Dirk Groeneveld and Jesse Dodge and Kyle Lo
+},
   year = {2024},
   journal={arXiv preprint},
 }
@@ -80,7 +84,7 @@ class Dolma(datasets.GeneratorBasedBuilder):
         for name in _URL_LISTS.keys()
     ]
 
-    DEFAULT_CONFIG_NAME = "v1_5r2"
+    DEFAULT_CONFIG_NAME = "v1_6"
 
     def _info(self):
         return datasets.DatasetInfo(
@@ -89,21 +93,25 @@ class Dolma(datasets.GeneratorBasedBuilder):
                 {
                     "id": datasets.Value("string"),
                     "text": datasets.Value("string"),
-                    "metadata": datasets.Value("string"),
+                    # "metadata": datasets.Value("string"),
                     "added": datasets.Value("string"),
-                    # "metadata": datasets.Value("")
+                    "created": datasets.Value("string"),
+                    "source": datasets.Value("string"),
                 }
             ),
             supervised_keys=None,
         )
 
-    def _split_generators(self, dl_manager):
-        with open(_URL_LISTS[self.config.name], mode="rt", encoding="utf-8") as f:
-            subset_urls = f.read().splitlines()
+    def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
+        path = dl_manager.download(_URL_LISTS[self.config.name])
 
-        breakpoint()
+        with open(path, mode="rt", encoding="utf-8") as f:  # type: ignore[no-untyped-call]
+            subset_urls = f.read().splitlines()
 
-        subset_files = dl_manager.download(subset_urls)
+        if _DATA_DIR is not None:
+            subset_files = [os.path.join(_DATA_DIR, url.replace(_BASE_URL, "").lstrip("/")) for url in subset_urls]
+        else:
+            subset_files = dl_manager.download(subset_urls)
 
         return [
             datasets.SplitGenerator(
@@ -112,6 +120,18 @@ class Dolma(datasets.GeneratorBasedBuilder):
             )
         ]
 
-    def _generate_examples(self, files):
+    def _generate_examples(self, files: List[str]):
         """This function returns the examples in the raw (text) form."""
-        raise NotImplementedError("Dolma is a streaming dataset")
+        for fn in files:
+            logger.info("generating examples from = %s", fn)
+
+            with gzip.open(fn, mode="rt", encoding="utf-8") as f:
+                for line in f:
+                    row = json.loads(line)
+                    yield row["id"], {
+                        "id": row["id"],
+                        "text": row["text"],
+                        "added": row.get("added", ""),
+                        "created": row.get("created", ""),
+                        "source": row.get("source", ""),
+                    }