| --- |
| language: |
| - ru |
| - en |
| license: cc0-1.0 |
| task_categories: |
| - text-generation |
| - conversational |
| tags: |
| - synthetic |
| - instruction-tuning |
| - bilingual |
| - russian |
| - english |
| - coding |
| - greetings |
| pretty_name: RU-EN Humanlike Synthetic Instructions |
| --- |
| |
| # RU-EN Humanlike Synthetic Instructions |
|
|
| Synthetic Russian + English prompt → answer dataset for SFT/chat fine-tuning experiments. |
|
|
| Current local build: **20,000 examples**. |
|
|
| ## What's new in v2 |
|
|
| Added another 10k examples and expanded the generator with: |
|
|
| - greetings and warm opening conversations; |
| - more small talk and human-like supportive replies; |
| - more coding categories; |
| - debugging prompts; |
| - code review prompts; |
| - tests and refactoring prompts; |
| - API design; |
| - data analysis; |
| - translation; |
| - summarization; |
| - comparison; |
| - polite emails; |
| - roleplay mentor dialogs; |
| - habit-building advice. |
|
|
| ## Files |
|
|
| - `data/train.jsonl` |
| - `data/validation.jsonl` |
| - `data/test.jsonl` |
| - `data/dataset_info.json` |
| - `generate_dataset.py` — reproducible generator; can create thousands or millions of rows. |
|
|
| ## Schema |
|
|
| Each row contains: |
|
|
| ```json |
| { |
| "id": "stable id", |
| "lang": "ru|en", |
| "category": "greeting|explain|planning|empathy|coding|coding_debug|...", |
| "prompt": "user prompt", |
| "response": "assistant response", |
| "messages": [ |
| {"role": "user", "content": "..."}, |
| {"role": "assistant", "content": "..."} |
| ], |
| "source": "synthetic_template_v2", |
| "license": "CC0-1.0", |
| "quality_note": "Template-generated; review/filter before serious training." |
| } |
| ``` |
|
|
| ## Generate more rows |
|
|
| ```bash |
| python generate_dataset.py --n 20000 --out data --seed 43 |
| python generate_dataset.py --n 1000000 --out data_million --seed 43 |
| ``` |
|
|
| ## Load with Hugging Face datasets |
|
|
| ```python |
| from datasets import load_dataset |
| |
| dataset = load_dataset( |
| "json", |
| data_files={ |
| "train": "data/train.jsonl", |
| "validation": "data/validation.jsonl", |
| "test": "data/test.jsonl", |
| } |
| ) |
| ``` |
|
|
| ## Intended use |
|
|
| Good for prototyping a Hugging Face dataset and testing SFT pipelines. For a production-quality model, mix this with reviewed high-quality examples, deduplicate, add preference data, and run safety/quality filters. |
|
|
| ## Design goals |
|
|
| - Russian + English coverage. |
| - Warm, natural, emotionally aware replies. |
| - Chat-friendly `messages` column. |
| - Expanded coding/task categories. |
| - Safer refusals for harmful requests. |
|
|
| ## Limitations |
|
|
| This is synthetic template-generated data. It can be repetitive at very large scale. Use it as a scaffold/base generator, not as the only source for a serious model. |
|
|
