IFEval-mini-120

A deterministic 120-prompt subset of google/IFEval that still covers all 25 instruction types. Built for evaluating small and compute-constrained models — especially temporally-gated / historical LMs, where a full 541-prompt sweep across many model variants is expensive.

Scoring is unchanged from IFEval, so per-category numbers remain interpretable; only the sample size differs.

Composition

prompts	120
instruction types covered	25 / 25
single-constraint prompts	69
two-constraint prompts	40
three-constraint prompts	11
total instruction instances	182

Selection is random.Random(0).shuffle(...) over the 541 IFEval rows, then the first 120 — fully reproducible via build_subset.py. Fields are byte-identical to the source rows (key, prompt, instruction_id_list, kwargs).

Usage

lm-evaluation-harness (scoring delegates to the harness's own IFEval implementation, so numbers are directly comparable to published IFEval results):

lm_eval --model hf --model_args pretrained=<model> \
        --tasks ifeval_mini120 --include_path ./ --output_path out/

Directly:

from datasets import load_dataset
ds = load_dataset("json", data_files={"test": "data/test.jsonl"})["test"]

Reference results

Measured on this exact subset — greedy decoding, seed 0, strict scoring. Useful as a sanity check that your harness is wired up correctly.

model	params	prompt-strict	inst-strict
SmolLM3-3B (/no_think)	3B	69.2	76.9
Mistral-7B-Instruct-v0.2	7B	49.2	59.3
typewriter-1938-7B-instruct	7B	17.5	27.5
typewriter-1913-7B-sft-v2	7B	15.8	23.6
typewriter-1913-7B-dpo-v2	7B	14.2	21.4
typewriter-1913-7B-base-v2	7B	14.2	23.6
violet-1b4-chat	1.4B	10.8	18.1
talkie-1930-13b-it (nf4)	13B	8.3	19.2

Scored through this task definition (seeded, deterministic). Talkie was run at bitsandbytes nf4 and batch=1; all others bf16 or fp8 via vLLM.

Caveats

• n=120 is small. The 95% Wilson interval on prompt-level accuracy is roughly ±7–9pp. Use this for iteration and for screening many variants; quote full IFEval for headline claims.
• Aggregate scores flatter base models. An untuned model that rambles satisfies length- and keyword-style constraints by accident, and can satisfy format constraints by echoing the prompt — IFEval's format prompts contain literal <<title>> examples. Report per-category, and consider reporting prompt/response n-gram overlap alongside any base-model score.
• Left-padded batching can silently corrupt some architectures. Verify your generations are not degenerate before trusting a low score.
• Upstream IFEval scoring is nondeterministic; this task fixes it. lm-eval-harness calls langdetect without seeding it, so language:response_language can flip between runs — we measured the same generations scoring 49.17 / 50.00 / 49.17 on three identical passes. utils.py here sets DetectorFactory.seed = 0, making repeated scoring reproducible. If you compare models within ~1pp on stock IFEval, some of that difference is noise.

License & attribution

Apache-2.0, inherited from google/IFEval (Zhou et al., 2023, Instruction-Following Evaluation for Large Language Models). This is a redistributed subset with no modification to prompts or scoring.

@article{zhou2023instructionfollowing,
  title={Instruction-Following Evaluation for Large Language Models},
  author={Zhou, Jeffrey and Lu, Tianjian and Mishra, Swaroop and Brahma, Siddhartha
          and Basu, Sujoy and Luan, Yi and Zhou, Denny and Hou, Le},
  journal={arXiv preprint arXiv:2311.07911},
  year={2023}
}