AmanPriyanshu/reasoning-sft-Nemotron-Instruction-Following-Chat-v1

---
license:
- cc-by-4.0
task_categories:
- text-generation
language:
- en
tags:
- reasoning
- sft
- chain-of-thought
- instruction-following
- structured-outputs
- chat
size_categories:
- 100K<n<1M
---

# Nemotron Instruction Following Chat v1 (Reasoning SFT)

Converted version of [nvidia/Nemotron-Instruction-Following-Chat-v1](https://huggingface.co/datasets/nvidia/Nemotron-Instruction-Following-Chat-v1), filtered to 157,595 rows where assistant responses include genuine reasoning traces (`reasoning_content`).

## Format

Each row has three columns:

- **`input`** — list of dicts with role/content conversation turns (system, user, and prior assistant turns up to the final assistant response)
- **`response`** — `<think>` block containing the model's reasoning followed by the final answer
- **`domain`** — task domain: `instruction_following` or `structured_outputs`

## Domain Distribution

| Domain | Rows |
|--------|------|
| instruction_following | 152,628 |
| structured_outputs | 4,967 |

## Conversion

- Source: both `chat_if` (426K rows) and `structured_outputs` (5K rows) splits
- Filtered to rows where the last assistant message contains non-empty `reasoning_content` (36.57% of total)
- Reasoning mapped into `<think>` blocks, answer follows after `</think>`
- Validated exactly 1 open and 1 close think tag per response
- Multi-turn conversations preserved: all prior turns (system, user, assistant) become input context

## Usage

```py
from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq
import random

repo = "AmanPriyanshu/reasoning-sft-Nemotron-Instruction-Following-Chat-v1"
path = hf_hub_download(repo_id=repo, filename="data.parquet", repo_type="dataset")
table = pq.read_table(path)
print(f"Loaded {len(table):,} rows\n")

i = random.randint(0, len(table) - 1)
row = {col: table.column(col)[i].as_py() for col in table.schema.names}
print(f"=== ROW (index {i}) ===")
print(f"\n[domain] {row['domain']}")
print(f"\n[input] ({len(row['input'])} turns)")
for t in row["input"]:
preview = t["content"][:300] + ("..." if len(t["content"]) > 300 else "")
print(f" {t['role']}: {preview}")
rp = row["response"][:1500]
if len(row["response"]) > 1500:
rp += "..."
print(f"\n[response]\n{rp}")
```

## License

Inherited from the [original dataset](https://huggingface.co/datasets/nvidia/Nemotron-Instruction-Following-Chat-v1) by NVIDIA Corporation.