ukisai/opsa_sft_50k_v1

This is a multi-turn SFT dataset with assistant turns included. It is intended for supervised fine-tuning, not OPSA prompt-only training.

The 50,000-row mix is:

• 15,209 Qwen3.8 multi-turn agent traces from ukisai/Qwen3.8-27B-multi-turn-agent-sft.
• 20,000 OpenThoughts-Agent traces from open-thoughts/OpenThoughts-Agent-SFT-100K, sampled 5,000 each from SWE-Smith, SuperUser, Tezos, and IssueTasks.
• 14,791 non-coding Nemotron SFT traces from the previous opsa_sft_50k_v1 build. This includes 7,315 available multi-turn Nemotron rows, plus 7,476 single-turn rows to reach the target size.

The result is 40,339 multi-turn conversations and 9,661 single-turn conversations. All rows end with an assistant turn.

Files

• data/opsa_sft_50k_v1.jsonl — full conversations with messages.
• data/train-00000-of-00001.parquet — the same rows, with messages_json.

Use the JSONL for normal SFT training and messages as the conversation field.

Row format

Each row is a full conversation:

{
  "row_id": "...",
  "source": "...",
  "source_path": "...",
  "source_row": 123,
  "source_seek": null,
  "row_sha1": "...",
  "messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}],
  "label": "",
  "source_task_id": "swesmith-12345_copy0000",
  "source_category": "swesmith",
  "source_trace_turns": 12,
  "trace_type": "full"
}

The Parquet file stores the same conversation as messages_json. OpenAI-style developer messages are normalized to system. A small number of Nemotron agentic rows also contain tool turns.

Provenance

Every row keeps its original source, source file, and row index.

• Qwen and OpenThoughts rows use absolute Parquet row indices. row_sha1 hashes the canonical JSON of the source row.
• Nemotron rows retain the original row_id, source_path, source_row, source_seek, and row_sha1 from the prior dataset build.

Licenses

• NVIDIA Nemotron portions: CC-BY-4.0.
• OpenThoughts-Agent-SFT-100K: Apache-2.0.
• Qwen3.8-27B-multi-turn-agent-sft: Apache-2.0.