You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access is reviewed manually by LANA-research. Please describe who you are and how you plan to use the data.
Log in or Sign Up to review the conditions and access this dataset content.
LANA-IF-Qwen3.8-Distill
Verified if supervised fine-tuning data distilled from Qwen3.8-27B by the LANA-research data engine, one config per data source (the source of the prompt). Every row is one teacher response that passed automatic verification, in the canonical slf-sft-v1 schema with full provenance.
Verification: every verifiable formatting constraint passes its checker (IFEval-style).
producer.stage/producer.modeidentify the sampling stage and thinking mode.
Subsets (one per source)
| config | source dataset | train | validation |
|---|---|---|---|
post_v2_chat |
nvidia/Nemotron-Post-Training-Dataset-v2:chat | 36,427 | 356 |
cascade_rl_if |
nvidia/Nemotron-Cascade-RL-Instruction-Following | 32,005 | 258 |
rl_instruction_following |
nvidia/Nemotron-RL-instruction_following | 17,339 | 185 |
post_v1_stem |
nvidia/Nemotron-Post-Training-Dataset-v1:stem | 12,311 | 109 |
if_chat_v1 |
nvidia/Nemotron-Instruction-Following-Chat-v1:chat_if | 9,984 | 106 |
post_v1_math |
nvidia/Nemotron-Post-Training-Dataset-v1:math | 6,207 | 51 |
Total: 115,338 rows. validation is a 1% hold-out by question_hash.
Loading
from datasets import load_dataset
ds = load_dataset("LANA-research/LANA-IF-Qwen3.8-Distill", "post_v2_chat", split="train")
row = ds[0]; prompt = row["messages_in"]; target = (row["reasoning"], row["response"])
Fields
Schema slf-sft-v1 (31 fields): identity (uid, family, family_version, domain, task_type, language); input (messages_in, tools, problem_text, question_hash, source_dataset, source_id, source_split, license, seed_pool, difficulty); reference (kind, answer, choices, tests_ref, constraints, gold_source); producer (teacher model, serving, mode, sampling, stage, sample index); gen_run (client, hash, time); target (reasoning = text inside <think>, response = visible answer, finish_reason, n_tokens); verify (method, version, passed, score, details); accepted (always true here), reject_reason, flags, decontam (method, checked_against, hit), split, meta (JSON).
Decontamination: exact + 13-gram checks of the prompt against MATH-500, GPQA-Diamond, IFEval, MMLU-Pro, LiveCodeBench v6, OJBench and other evaluation inputs (per-row record in decontam).
License
Mixed: each row's license field gives the license of its prompt source; responses were generated with Qwen3.8-27B.
- Downloads last month
- 35