locailabs/ultrachat_nemotron_120b

---
language:
- en
license: cc-by-4.0
task_categories:
- text-generation
tags:
- ultrachat
- synthetic
- nemotron
---

# UltraChat Responses (Nemotron-3-Super-120B)

Synthetic general-purpose chat responses generated from UltraChat prompts using
Nemotron-3-Super-120B, with the Jupiter (Locai Labs) system prompt.

## Splits

| Split | Count | Description |
|-------|-------|-------------|
| `no_reasoning` | 5822 | Reasoning disabled — direct responses |
| `reasoning` | 2376 | Reasoning enabled — assistant content prefixed with `<think>...</think>` |

## How this dataset was made

### 1. Prompt sourcing

Prompts were sampled from UltraChat, a large-scale multi-turn instruction dataset.
A subset of 10k single-turn prompts was used.

### 2. Response generation

Responses were generated using **Nemotron-3-Super-120B** via a self-hosted vLLM
endpoint, paired with the Jupiter / Locai Labs system prompt (not included in the
output messages).

Two variants were generated:
- **no_reasoning**: `enable_thinking: false` in the chat template
- **reasoning**: default (thinking enabled), with the model's reasoning trace
wrapped in `<think>...</think>` tags prepended to the assistant message

### 3. Format

Each row contains a `messages` list in standard chat format:

```json
[
{"role": "user", "content": "..."},
{"role": "assistant", "content": "..."}
]
```

For the `reasoning` split, assistant content looks like:

```
<think>
...reasoning trace...
</think>
...final answer...
```

## Intended use

Fine-tuning / post-training LLMs for general instruction-following with the
Jupiter model identity.