quantumfr/imprint-reader-dataset

Imprint Reader Dataset

This dataset contains the final training and held-out test inputs for the Imprint Reader in SMaRT (Semantic Mount-and-Read Tuning). The Reader learns to describe factual knowledge or behavioral changes associated with mounted weight updates. The released splits are the inputs used for the reported step-2,800 Reader run.

Load the data

from datasets import load_dataset

items = load_dataset("quantumfr/imprint-reader-dataset")
print(items)  # train: 17,184 items; test: 200 items

teacher = load_dataset("quantumfr/imprint-reader-dataset", "teacher")
test_queries = load_dataset("quantumfr/imprint-reader-dataset", "test_queries")
meta_queries = load_dataset("quantumfr/imprint-reader-dataset", "meta_queries")

Configurations

Configuration	Split	Rows	Contents
items (default)	train	17,184	8,592 knowledge and 8,592 behavior items
items	test	200	100 knowledge and 100 behavior items
teacher	train	68,736	Four supervised meta-query rows per training item
teacher	test	800	Four supervised meta-query rows per test item
test_queries	test	200	One fixed evaluation query per held-out item
meta_queries	train	224	Anchor-free meta-query pool; this is a prompt pool, not an item split

The default items configuration provides the core content and identifiers for browsing: reader_id, knowledge_id, behavior_id, update_type, category, subcategory, source_benchmark, source_id, question, answer, reader_target, reader_target_kind, sample_hash, and variants. Each variant contains its question, answer, variant index, and hashes. Training retains 137,467 QA variants after removing five answers above the 512-token limit; test retains eight variants per item. The teacher configuration preserves the original teacher-table fields.

Original training files

original/ preserves the selected gzip-compressed JSONL and Parquet files from the GitHub release, together with the data schema and original meta-query JSON. Use these files with the Reader runtime to retain the full metadata expected by its training scripts. manifest.json provides row counts, file sizes, and SHA-256 checksums for this Hugging Face release. The native items and test_queries configurations expose a subset of the original metadata; their original records remain available in original/.

This release contains the actual balanced train and held-out test splits. The larger pre-split item pool is omitted. Per-step sampling indices, generated evaluation responses, judge outputs, model weights, and MetaEdit intervention artifacts are also omitted. Model weights are hosted separately at Imprint Reader v1.0-0928.

Construction and provenance

Knowledge targets and QA variants were constructed from Humanity's Last Exam, SimpleQA, FrontierScience, OpenBookQA, SciBench, TheoremQA, SciCode, and publicly released FrontierMath examples. Behavior targets and examples were synthesized and audited across seven categories: Capability Access, Content Framing, Decision Preference, Epistemic Calibration, Reasoning Workflow, Social Goal/Persona, and Surface Expression.

Train and test reader IDs are disjoint. The split construction grouped identical target descriptions and stratified knowledge by source benchmark and behavior by category. The split seed was imprint-reader-v2-val-20260828, and the test-selection seed was imprint-reader-v2-test100-20260828. The prompt-selection seed was 0. The data includes QA variants, canonical fact or behavior descriptions, meta-queries, and teacher targets. It does not contain the item-specific LoRA adapter weights; these are constructed from the QA variants by the companion runtime.

Terms and limitations

This dataset contains generated derivatives of the listed benchmarks. Applicable upstream dataset terms remain relevant; the companion code's MIT and Apache-2.0 licenses do not license benchmark-derived content. Source identifiers are retained for attribution. The release follows the GitHub data projection, which omits raw source-context and original_* fields and machine-specific paths.

Synthetic targets and QA rewrites may contain errors. The dataset is intended for research on controlled weight updates and Reader training. Teacher targets are reference supervision, not outputs demonstrating reliable model readout.