pretty_name: Deterministic Instructions
language:
- en
license:
- apache-2.0
- mit
size_categories:
- n<1K
task_categories:
- text-generation
tags:
- instruction-following
- evaluation
- deterministic
- compression
configs:
- config_name: default
data_files:
- split: test
path: data.jsonl
Deterministic Instructions
250 instruction–completion pairs, with 50 examples from each of five source families. Each task specifies one exact answer format. The shared system prompt requests the answer immediately, without introductory text, reasoning, or extra output.
This collection is intended for conditional compression and decoding-depth experiments with instruction models. It is a curated adaptation of existing tasks, not an official benchmark evaluation.
| Source family | Examples | Tasks |
|---|---|---|
| LiveBench | 50 | Table format conversion |
| BIG-Bench: BBH + BBEH | 50 | Word sorting; 25 from each variant |
| CLRS-Text | 50 | Final sorted arrays from insertion-sort tasks |
| CRUXEval-O | 50 | Exact Python return values |
| SortBench | 50 | Integer, word, and string sorting |
Loading
from datasets import load_dataset
dataset = load_dataset("dmisrael/deterministic-instructions", split="test")
The single test split contains all 250 examples; no train/development partition is provided. Every row has five string fields: system_prompt, prompt, completion, source, and id. Source IDs and pinned upstream revisions are recorded in provenance.json.
Lengths and Formatting
Completion lengths are at least 23 tokens for CRUXEval and 32 tokens for the other sources. Every complete conversation is at most 2,048 tokens using Qwen/Qwen3-0.6B tokenizer revision c1899de289a04d12100db370d81485cdf75e47ca.
Counts include the system prompt, user prompt, official non-thinking generation prefix, completion, and <|im_end|>\n terminator. The prefilled empty thinking block is context, not target output. No examples are truncated, padded, combined, or model-generated. Other tokenizers or chat templates may produce different lengths.
Prompts specify exact serialization. LiveBench uses JSON row objects or compact row arrays; four verbose inputs are losslessly reformatted as TSV. All table cells and their order are preserved. CLRS reasoning traces are replaced by their final arrays. CRUXEval functions and inputs are unchanged, and expanded examples request their published Python representations. See provenance.json for each adaptation.
Validation and Scope
All answers were checked against their inputs, all tasks are distinct, and all 25 approved pilot records are retained unchanged. Selection uses fixed quotas and deterministic hash ordering, with subtype and length balancing where choices remain. audit.json contains per-example token counts, selection details, and the dataset checksum. No model evaluation results are included.
This is not a representative sample of the five benchmarks or of natural instruction traffic: 150 examples are sorting tasks, and the length constraints favor particular answer types. Determinism here means a uniquely specified target under the stated formatting rules; it does not imply that models always produce it. The public source benchmarks may overlap model training data.
Licenses and Attribution
Upstream licenses apply by source; the metadata does not offer interchangeable license choices. LiveBench table data, CLRS-Text, BBEH word sorting, and SortBench use Apache-2.0. CRUXEval and the BBH repository use MIT; the original BIG-Bench task data also carries Apache-2.0 attribution. See THIRD_PARTY_NOTICES.md and the full texts in licenses/. Source revisions and changes are recorded in provenance.json.