Datasets:

willhx
/

if_oracle_sft

license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - instruction-following
  - ifeval
  - sft
  - rejection-sampling
pretty_name: IF Oracle SFT
configs:
  - config_name: default
    data_files:
      - split: train
        path: if_oracle_sft-*.parquet

IF Oracle SFT

Oracle instruction-following SFT data built by rejection sampling from willhx/Qwen3-8B-Base-IF.

How it was made

Policy: willhx/Qwen3-8B-Base-IF (a trained IF RL checkpoint).
Prompts: the full allenai/IF_multi_constraints_upto5-derived training set (IF_multi_constraints_upto5_ifbench_en, 88,832 prompts covered).
Sampling: 8 responses per prompt, temperature 1.0, max 8192 response tokens (rollout via slime + SGLang, --debug-rollout-only).
Reward: rule-based IFEval-G (rm_type=multi) — the fraction of the prompt's instruction constraints that a response satisfies.
Oracle filter: only responses with reward == 1.0 (ALL constraints satisfied) are kept. A prompt may contribute multiple oracle responses.

Columns

column	description
`prompt`	user instruction (clean text)
`response`	model response that satisfies all constraints
`messages`	`[{user}, {assistant}]` chat form of the pair
`reward`	always `1.0` (oracle)
`instruction_id_list`	IFEval-G instruction ids the prompt imposes
`kwargs`	JSON string of per-instruction arguments
`source`, `record_id`, `rm_type`	provenance from the source dataset
`rollout_id`	which rollout batch the sample came from

Load

from datasets import load_dataset
ds = load_dataset("willhx/if_oracle_sft", split="train")