---
language:
- en
license: other
task_categories:
- text-generation
tags:
- chat
- conversational
- sft
- warm-tone
pretty_name: warm-chat-12k
size_categories:
- 10K<n<100K
---

# warm-chat-12k

11,400 warm conversational SFT rows + 600 held-out eval rows, `messages` format
(`role`/`content`). Built because every modern public chat set is either a
technical manual or a roleplay convention, and neither teaches a model to just...
talk like a person.

Filtered for genuine chat: no code, no math, no fiction, no roleplay, no
refusals, no "as an AI" disclaimers, no markdown lectures. What's left is
greetings, advice, everyday questions, and casual back-and-forth — with feeling.

## Composition

| Source | Rows | Teacher |
|---|---|---|
| Verdugie/opus-candid-training-data | 1,136 | Claude Opus 4.6 |
| nvidia/Nemotron-SFT-Instruction-Following-Chat-v2 (reasoning_off) | 10,264 | Kimi-K2 / GLM-4.6 / Qwen3-235B-Instruct |

Every row carries a `src` field with its origin. License note: Nemotron data is
ODC-By (attribution required — this card is that attribution). Check the
opus-candid repo for its terms before commercial use.

## Example

user: hello

assistant: Hey! ٩(◕‿◕｡)۶ How's it going? I'm here to chat or help with whatever you need — just let me know!

## Use

```python
from datasets import load_dataset
ds = load_dataset("minas2025/warm-chat-12k")