ukisai/opsa_50k_training_v1

48,811 deduplicated, shuffled training prompts for OPSA (On-Policy Self-Adaptation, arXiv:2608.31046) on the swifter DeepSeek-V4.1-Flash checkpoint. Built from NVIDIA's Nemotron post-training datasets (the post-training data of Nemotron Nano/Super/Ultra v3, Super 3.5 VL and Lightning 3.5, refreshed Sep-Oct 2026) as a fresher, multi-domain replacement for single-scaffold agent SFT traces.

Prompts only: OPSA generates fresh responses on-policy; dataset responses are never trained on. Source licenses are CC BY 4.0 / Apache 2.0 / NVIDIA open model data licenses; this derived prompt set is distributed under CC BY 4.0.

Format

• opsa_50k_training_v1.jsonl - slime-ready rows: {"prompt": [{"role", "content"}...], "label": "", "source": "nvidia/...", "tools": [...]} (tools present only when the source provides function definitions; use with slime --tool-key tools if desired).
• train-00000-of-00001.parquet - same rows with tools serialized as tools_json (string).

Provenance (v1.1)

Every row records where it came from, verifiable against the source repo:

Field	Meaning
source	source dataset repo
source_path	file inside the source repo
source_row	JSONL: 0-based line index after seeking to source_seek bytes and discarding the partial boundary line. Parquet: absolute row index in the shard
source_seek	byte offset the reader seeked to (JSONL only; parquet rows have null)
row_sha1	JSONL: sha1 of the exact source line bytes (trailing newline stripped). Parquet: sha1 of json.dumps(row, sort_keys=True, ensure_ascii=False)
row_id	sha1 of `source

Verification recipe (JSONL rows):

from huggingface_hub import HfFileSystem
import hashlib
fs = HfFileSystem()
with fs.open(f"datasets/{row['source']}/{row['source_path']}") as f:
    f.seek(row["source_seek"]); f.readline()  # discard partial boundary line
    for _ in range(row["source_row"]):
        f.readline()
    line = f.readline().strip()
assert hashlib.sha1(line).hexdigest() == row["row_sha1"]

Composition

Source	Rows	Extracted as
nvidia/Nemotron-SFT-Math-v4	12,000	problem (label = expected answer)
nvidia/Nemotron-SFT-Science-v2	6,000	system + first user turn
nvidia/Nemotron-SFT-Instruction-Following-Chat-v3	4,471	system + first user turn
nvidia/Nemotron-SFT-OpenCode-v1	4,448	agent_prompt
nvidia/Nemotron-SFT-ARC-AGI-v1	3,746	system + first user turn
nvidia/Nemotron-SFT-Multilingual-v2	3,000	system + first user turn
nvidia/Nemotron-SFT-SWE-v3	1,637	system + first user turn
nvidia/Nemotron-SFT-Competitive-Programming-v2	1,501	system + first user turn
nvidia/Nemotron-RL-Instruction-Following-Structured-Outputs-v2	1,500	Responses-API input
nvidia/Nemotron-SFT-Safety-v1	1,500	system + first user turn
nvidia/Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1	1,500	Responses-API input + tools
nvidia/Nemotron-SFT-Agentic-v2	1,200	system + first user turn
nvidia/Nemotron-RL-Agentic-Function-Calling-Pivot-v1	1,000	Responses-API input + tools
nvidia/Nemotron-RL-Science-v1	1,000	problem (label = expected answer)
nvidia/Nemotron-RL-Safety-v1	945	prompt
nvidia/Nemotron-RL-Identity-Following-v1	500	Responses-API input
nvidia/Nemotron-RL-litmus-bench-v0.1	500	Responses-API input
nvidia/Nemotron-RL-Instruction-Following-Free-Form-Formatting-v1	500	Responses-API input
nvidia/Nemotron-RL-Instruction-Following-Citation-Formatting-v1	500	Responses-API input
nvidia/Nemotron-RL-Instruction-Following-Calendar-v2	500	Responses-API input
nvidia/Nemotron-RL-Agentic-SWE-Pivot-v1	500	Responses-API input + tools

Build

Regenerate with the UkisAI/opsa-run tool tools/prepare_nemotron_prompts.py (seed 1234, dedup by prompt hash, 30k-char prompt cap, thin streamed slices of the large shards). Domain weighting follows the OPSA paper's long-form reasoning mechanism: math + code + agentic + science dominate; short chat/IF prompts are a minority.