---
language:
- ru
license: cc-by-nc-sa-4.0
task_categories:
- text-generation
tags:
- russian
- colloquial
- style-transfer
- sft
- gemma
size_categories:
- 10K<n<100K
---

# Russian colloquial SFT (50k, mat-free)

**English:** ~50,000 chat SFT examples: rewrite formal Russian into casual Telegram-style Russian **without profanity**.

## Format

Each line is JSON with a `messages` array (ShareGPT / TRL):

```json
{
  "messages": [
    {
      "role": "user",
      "content": "Перепиши простым разговорным русским, как в переписке. Без мата и грубости. Сохрани смысл:\n<формальный текст>"
    },
    {
      "role": "assistant",
      "content": "<разговорный ответ>"
    }
  ]
}
```

Files in this repo:

| File | Description |
|------|-------------|
| `russian_colloquial_sft.jsonl` | Training split (50,000 rows) |
| `russian_colloquial_sft.meta.jsonl` | Metadata: `source_id`, HF row, licenses, lengths |

## How it was built

1. **Source pool:** informal Russian lines from [kurumikz/telegram-corpus-russian-kazakh](https://huggingface.co/datasets/kurumikz/telegram-corpus-russian-kazakh) and related social text (filtered, deduplicated, profanity removed from candidates).
2. **Pair generation:** batched **Gemini** (`gemini-2.0-flash-lite`) produced `{formal, colloquial}` per line.
3. **Validation:** pairs with profanity in `formal` or `colloquial` were dropped.
4. **Training fields:** `formal` → user prompt (with fixed instruction prefix); `colloquial` → assistant.

## Licenses

- Generated pairs: use under your project terms when publishing derivatives.
- Upstream Telegram corpus: **CC-BY-NC-SA-4.0** (see `license` in `*.meta.jsonl`).
- Do not use for commercial training without checking upstream licenses.

## Used to train

| Model | Hub |
|-------|-----|
| LoRA adapter | [pavelfedortsov/gemma4-e4b-lora-colloquial-ru](https://huggingface.co/pavelfedortsov/gemma4-e4b-lora-colloquial-ru) |
| Merged weights | [pavelfedortsov/gemma4-e4b-colloquial-ru-merged](https://huggingface.co/pavelfedortsov/gemma4-e4b-colloquial-ru-merged) |

Training config: `config/train_colloquial_e4b_gpu.yaml` in the gemma4 fine-tuning project (same layout as the model repos on Hub).