|
| 1 | +--- |
| 2 | +date: 2026-04-16 |
| 3 | +authors: |
| 4 | + - nmulepati |
| 5 | + - davanstrien |
| 6 | +--- |
| 7 | + |
| 8 | +# **Push Datasets to Hugging Face Hub** |
| 9 | + |
| 10 | +You just generated 10k multilingual greetings (or some other cool dataset). Now what — email a parquet file? |
| 11 | +Nah. Call `.push_to_hub()` and you've got a live dataset page on Hugging Face. Done and dusted 🚢. |
| 12 | + |
| 13 | +<!-- more --> |
| 14 | + |
| 15 | +{ width=100% } |
| 16 | + |
| 17 | +--- |
| 18 | + |
| 19 | +Here's the full flow — build a multilingual greeting dataset with a conversation |
| 20 | +training processor, generate it, and push it to the Hub in one go: |
| 21 | + |
| 22 | +```python |
| 23 | +import data_designer.config as dd |
| 24 | +from data_designer.interface import DataDesigner |
| 25 | + |
| 26 | +data_designer = DataDesigner() |
| 27 | +config_builder = dd.DataDesignerConfigBuilder() |
| 28 | + |
| 29 | +config_builder.add_column( |
| 30 | + dd.SamplerColumnConfig( |
| 31 | + name="language", |
| 32 | + sampler_type=dd.SamplerType.CATEGORY, |
| 33 | + params=dd.CategorySamplerParams( |
| 34 | + values=["English", "Spanish", "French", "German", "Italian"], |
| 35 | + ), |
| 36 | + drop=True, |
| 37 | + ) |
| 38 | +) |
| 39 | + |
| 40 | +config_builder.add_column( |
| 41 | + dd.LLMTextColumnConfig( |
| 42 | + name="greeting", |
| 43 | + model_alias="nvidia-text", |
| 44 | + prompt="Write a casual greeting in {{ language }}.", |
| 45 | + ) |
| 46 | +) |
| 47 | +config_builder.add_column( |
| 48 | + dd.LLMTextColumnConfig( |
| 49 | + name="response", |
| 50 | + model_alias="nvidia-text", |
| 51 | + prompt="Write a helpful agent response to this greeting: '{{ greeting }}'.", |
| 52 | + ) |
| 53 | +) |
| 54 | + |
| 55 | +# Reshape into an OpenAI-style conversation training format |
| 56 | +config_builder.add_processor( |
| 57 | + dd.SchemaTransformProcessorConfig( |
| 58 | + name="conversations", |
| 59 | + template={ |
| 60 | + "messages": [ |
| 61 | + {"role": "user", "content": "{{ greeting }}"}, |
| 62 | + {"role": "assistant", "content": "{{ response }}"}, |
| 63 | + ] |
| 64 | + }, |
| 65 | + ) |
| 66 | +) |
| 67 | + |
| 68 | +results = data_designer.create(config_builder, num_records=10_000) |
| 69 | + |
| 70 | +# Ship it: |
| 71 | +url = results.push_to_hub( |
| 72 | + "my-org/multilingual-greetings", |
| 73 | + "10k synthetic agent/user conversations across 5 languages.", |
| 74 | + tags=["greetings", "multilingual", "conversation"], |
| 75 | +) |
| 76 | +print(url) # https://huggingface.co/datasets/my-org/multilingual-greetings |
| 77 | +``` |
| 78 | + |
| 79 | +--- |
| 80 | +## Two Ways In - same outcome |
| 81 | + |
| 82 | +**From results** (the happy path) — you just ran `.create()`, you have the |
| 83 | +results object, call `.push_to_hub()` on it. |
| 84 | + |
| 85 | +**From a folder** (the "I closed my notebook" path) — you saved artifacts to |
| 86 | +disk earlier and want to push them later: |
| 87 | + |
| 88 | +```python |
| 89 | +from data_designer.integrations.huggingface import HuggingFaceHubClient |
| 90 | + |
| 91 | +url = HuggingFaceHubClient.push_to_hub_from_folder( |
| 92 | + dataset_path="./my-saved-dataset", |
| 93 | + repo_id="my-org/multilingual-greetings", |
| 94 | + description="10k synthetic agent/user conversations across 5 languages.", |
| 95 | +) |
| 96 | +``` |
| 97 | + |
| 98 | +--- |
| 99 | +## What You Get on the Hub |
| 100 | + |
| 101 | +Once pushed, your dataset is live in the Hugging Face ecosystem: |
| 102 | + |
| 103 | +- **Dataset Viewer** — browsable in the browser immediately. Each processor |
| 104 | + config shows up as a separate subset tab (more on this in |
| 105 | + [Processors Get First-Class Treatment](#processors-get-first-class-treatment)). |
| 106 | +- **Streaming** — parquet means consumers can stream without downloading: |
| 107 | + |
| 108 | + ```python |
| 109 | + from datasets import load_dataset |
| 110 | + |
| 111 | + ds = load_dataset("my-org/multilingual-greetings", "conversations", split="train", streaming=True) |
| 112 | + ``` |
| 113 | + |
| 114 | +- **[Dataset Viewer API](https://huggingface.co/docs/dataset-viewer/)** — row |
| 115 | + pagination, text search, column statistics, and parquet shard URLs with no |
| 116 | + extra setup. |
| 117 | + |
| 118 | +--- |
| 119 | +## What Gets Uploaded |
| 120 | + |
| 121 | + |
| 122 | + |
| 123 | +Everything. The upload pipeline runs in this order: |
| 124 | + |
| 125 | +``` |
| 126 | +1. README.md ← auto-generated dataset card |
| 127 | +2. data/*.parquet ← your main dataset (remapped from parquet-files/) |
| 128 | +3. images/* ← if you have image columns (skipped otherwise) |
| 129 | +4. {processor}/* ← processor outputs (remapped from processors-files/) |
| 130 | +5. builder_config.json |
| 131 | +6. metadata.json ← paths rewritten to match HF repo layout |
| 132 | +``` |
| 133 | +
|
| 134 | +Each step is its own commit on the HF repo, so you get a clean history. |
| 135 | +
|
| 136 | +This is especially nice for large datasets. Data Designer writes output in |
| 137 | +batched parquet partitions — generate 100k records and you'll have dozens of |
| 138 | +parquet files across `parquet-files/`, `processors-files/`, and maybe `images/`. |
| 139 | +Manually uploading all of that, organizing it into the right HF repo structure, |
| 140 | +writing the dataset card YAML configs, and rewriting metadata paths would be |
| 141 | +tedious and error-prone. `push_to_hub` handles the whole thing in one call — |
| 142 | +folder uploads, path remapping, config registration, dataset card generation, |
| 143 | +all of it. |
| 144 | +
|
| 145 | +Re-pushing to the same `repo_id` updates the existing repo — no need to delete |
| 146 | +and recreate. |
| 147 | +
|
| 148 | +--- |
| 149 | +## Processors Get First-Class Treatment |
| 150 | +
|
| 151 | + |
| 152 | +
|
| 153 | +Notice the `SchemaTransformProcessorConfig` in the example above. That's doing |
| 154 | +the heavy lifting — it takes the raw `greeting` and `response` columns and |
| 155 | +reshapes each row into an OpenAI-style `messages` array: |
| 156 | +
|
| 157 | +```python |
| 158 | +config_builder.add_processor( |
| 159 | + dd.SchemaTransformProcessorConfig( |
| 160 | + name="conversations", |
| 161 | + template={ |
| 162 | + "messages": [ |
| 163 | + {"role": "user", "content": "{{ greeting }}"}, |
| 164 | + {"role": "assistant", "content": "{{ response }}"}, |
| 165 | + ] |
| 166 | + }, |
| 167 | + ) |
| 168 | +) |
| 169 | +``` |
| 170 | + |
| 171 | +The template is Jinja2 all the way down. Keys become columns in the output, |
| 172 | +values get rendered per-row with the actual column data. The template dict must |
| 173 | +be JSON-serializable — strings, lists, nested objects, all fair game. So you can |
| 174 | +build arbitrarily complex conversation schemas (multi-turn, system prompts, |
| 175 | +tool calls) just by adding more entries to the `messages` list. |
| 176 | + |
| 177 | +The processor runs after each batch and writes its output to a separate parquet |
| 178 | +file alongside the main dataset. The main dataset (`data/`) still has the raw |
| 179 | +columns — the processor output is an *additional* view, not a replacement. |
| 180 | + |
| 181 | +**When you push to hub, each processor gets its own top-level directory and its |
| 182 | +own HF dataset config.** So the `conversations` processor from our example ends |
| 183 | +up like this on HF: |
| 184 | + |
| 185 | +``` |
| 186 | +my-org/multilingual-greetings/ |
| 187 | +├── README.md |
| 188 | +├── data/ |
| 189 | +│ ├── batch_00000.parquet ← raw columns (greeting, response) |
| 190 | +│ └── batch_00001.parquet |
| 191 | +├── conversations/ |
| 192 | +│ ├── batch_00000.parquet ← transformed (messages array) |
| 193 | +│ └── batch_00001.parquet |
| 194 | +├── builder_config.json |
| 195 | +└── metadata.json |
| 196 | +``` |
| 197 | + |
| 198 | +The dataset card YAML frontmatter registers each processor as its own named |
| 199 | +config: |
| 200 | + |
| 201 | +```yaml |
| 202 | +configs: |
| 203 | +- config_name: data |
| 204 | + data_files: "data/*.parquet" |
| 205 | + default: true |
| 206 | +- config_name: conversations |
| 207 | + data_files: "conversations/*.parquet" |
| 208 | +``` |
| 209 | +
|
| 210 | +So consumers grab exactly the format they need: |
| 211 | +
|
| 212 | +```python |
| 213 | +from datasets import load_dataset |
| 214 | + |
| 215 | +# Raw columns — good for analysis |
| 216 | +df = load_dataset("my-org/multilingual-greetings", "data", split="train") |
| 217 | + |
| 218 | +# Conversation format — ready for fine-tuning |
| 219 | +df_conv = load_dataset("my-org/multilingual-greetings", "conversations", split="train") |
| 220 | +print(df_conv[0]) |
| 221 | +# {'messages': [{'role': 'user', 'content': 'Hey! Como estás?'}, |
| 222 | +# {'role': 'assistant', 'content': 'Hola! Estoy bien, gracias...'}]} |
| 223 | +``` |
| 224 | + |
| 225 | +The Quick Start section in the generated README includes these snippets |
| 226 | +automatically — one `load_dataset` call per processor. |
| 227 | + |
| 228 | +**Metadata paths are rewritten too.** Local paths like |
| 229 | +`processors-files/conversations/batch_00000.parquet` become |
| 230 | +`conversations/batch_00000.parquet` so file references in the metadata match |
| 231 | +the actual HF repo structure. |
| 232 | + |
| 233 | +If there are no processors, all of this is silently skipped — no empty |
| 234 | +directories, no phantom configs. |
| 235 | + |
| 236 | +--- |
| 237 | +## The Auto-Generated Dataset Card |
| 238 | + |
| 239 | +This is the fun part. The upload generates a full HuggingFace dataset card from |
| 240 | +your run metadata. It pulls from `metadata.json` and `builder_config.json` to |
| 241 | +build: |
| 242 | + |
| 243 | +- A **Quick Start** section with `load_dataset` code (including processor subsets) |
| 244 | +- A **Dataset Summary** with record count, column count, completion % |
| 245 | +- A **Schema & Statistics** table — per-column type, uniqueness, null rate, token stats |
| 246 | +- **Generation Details** — how many columns of each config type |
| 247 | +- A **Citation** block so people can cite your dataset |
| 248 | + |
| 249 | +Tags default to `["synthetic", "datadesigner"]` plus whatever you pass in. |
| 250 | +Size category (`n<1K`, `1K<n<10K`, etc.) is auto-computed. These tags make your |
| 251 | +dataset discoverable in [Hub search](https://huggingface.co/datasets?library=datadesigner&sort=trending) |
| 252 | +— you can browse all Data Designer datasets in one place. |
| 253 | + |
| 254 | +The template lives at `packages/data-designer/src/data_designer/integrations/huggingface/dataset_card_template.md` if you |
| 255 | +want to see the Jinja2 source. |
| 256 | + |
| 257 | +--- |
| 258 | +## Auth |
| 259 | + |
| 260 | +Token resolution follows the standard `huggingface_hub` chain: |
| 261 | + |
| 262 | +1. Explicit `token=` parameter |
| 263 | +2. `HF_TOKEN` env var |
| 264 | +3. Cached creds from `hf auth login` |
| 265 | + |
| 266 | +If none of those work, you get a clear error telling you what to do. |
| 267 | + |
| 268 | +--- |
| 269 | +## Reproducible Pipelines — The Round-Trip |
| 270 | + |
| 271 | +{ width="800" } |
| 272 | + |
| 273 | +Here's the payoff: every dataset you push includes `builder_config.json` — the |
| 274 | +full SDG pipeline definition. Anyone (including future-you) can recreate the |
| 275 | +exact same pipeline from the HuggingFace URL: |
| 276 | + |
| 277 | +```python |
| 278 | +import data_designer.config as dd |
| 279 | + |
| 280 | +config_builder = dd.DataDesignerConfigBuilder.from_config( |
| 281 | + "https://huggingface.co/datasets/my-org/multilingual-greetings/blob/main/builder_config.json" |
| 282 | +) |
| 283 | +``` |
| 284 | + |
| 285 | +That's it. One line. `from_config` accepts a raw URL, a local file path, a dict, |
| 286 | +or a YAML string. When you hand it a HuggingFace Hub URL, it auto-rewrites the |
| 287 | +blob URL to a raw URL behind the scenes so the fetch just works (same trick for |
| 288 | +GitHub blob URLs). |
| 289 | + |
| 290 | +The loaded config builder comes back fully hydrated — columns, model configs, |
| 291 | +constraints, seed config, all of it. You can inspect it, tweak it, and re-run: |
| 292 | + |
| 293 | +```python |
| 294 | +from data_designer.interface import DataDesigner |
| 295 | + |
| 296 | +# Maybe bump the count or swap a model |
| 297 | +results = DataDesigner().create(config_builder, num_records=50_000) |
| 298 | + |
| 299 | +# And push the new version right back |
| 300 | +results.push_to_hub( |
| 301 | + "my-org/multilingual-greetings-v2", |
| 302 | + "50k version with the same pipeline.", |
| 303 | +) |
| 304 | +``` |
| 305 | + |
| 306 | +So the full loop is: **design → generate → push → share URL → recreate → iterate**. |
| 307 | +The `builder_config.json` on HuggingFace *is* the reproducibility artifact. |
| 308 | + |
| 309 | +--- |
| 310 | +## Gotchas |
| 311 | + |
| 312 | +- **`repo_id` must be `username/dataset-name`** — exactly one slash. The client |
| 313 | + validates this before hitting the network. |
| 314 | +- **`description` is required** — it's the prose that appears right under the |
| 315 | + title on the dataset card. Make it good. |
| 316 | +- **`private=True`** if you don't want the world to see your dataset yet. You |
| 317 | + can flip it to public later from the dataset settings page. |
| 318 | +- **Metadata paths get rewritten** — local paths like `parquet-files/batch_00000.parquet` |
| 319 | + become `data/batch_00000.parquet` in the uploaded `metadata.json` so references |
| 320 | + stay valid on HF. |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?v=4&size=40){kind=avatar}
0 commit comments