reasoning-sft-Nemotron-Cascade-SFT-SWE-210K

Converted version of nvidia/Nemotron-Cascade-SFT-SWE, filtered to thinking=True rows with exactly one valid <think>...</think> block.

Format

Each row has three columns:

• input — list of dicts (conversation turns with role and content, system messages dropped, last assistant message removed)
• response — assistant response string including <think> reasoning block
• domain — {category}_{source} (e.g. SWE Repair_SWE-Fixer-Train)

Stats

• Original rows: ~214K
• Converted (thinking=True, clean think tags): ~210K

Usage

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

fpath = hf_hub_download(
    repo_id="AmanPriyanshu/reasoning-sft-Nemotron-Cascade-SFT-SWE-210K",
    repo_type="dataset",
    filename="data_converted.parquet",
    local_dir="./tmp_swe_peek"
)

pf = pq.ParquetFile(fpath)
rows = {"input": [], "response": [], "domain": []}
for batch in pf.iter_batches(batch_size=65_536):
    d = batch.to_pydict()
    rows["input"].extend(d["input"])
    rows["response"].extend(d["response"])
    rows["domain"].extend(d["domain"])

total = len(rows["input"])
for idx in random.sample(range(total), 3):
    print(f"\n{'='*80}\nRow {idx:,} / {total:,}  |  domain: {rows['domain'][idx]}\n{'='*80}")
    for msg in rows["input"][idx]:
        print(f"\n  [{msg['role']}]\n  {msg['content'][:300]}{'...' if len(msg['content']) > 300 else ''}")
    print(f"\n[response]\n{rows['response'][idx][:600]}{'...' if len(rows['response'][idx]) > 600 else ''}")

License

CC BY 4.0

Credits

Original dataset: nvidia/Nemotron-Cascade-SFT-SWE by NVIDIA
Responses generated by DeepSeek-R1-0528