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

Nemotron Instruction Following Chat v1 (Reasoning SFT)

Converted version of 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

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

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 by NVIDIA Corporation.