lamm-mit/gemma4-materials-mechanism-prompts

Gemma 4 Materials-Mechanism Prompt Corpus

This dataset collects the exact scientific prompts and registered prompt metadata used in “Reading and Steering Materials Science-Mechanism Representations in an Open-Weight Language Model” by Markus J. Buehler. It is organized as 21 Hugging Face configurations so that historical development prompts, frozen evaluations, falsification tests, and exploratory follow-ups are not pooled into one ambiguous table.

The release is a prompt and protocol resource. It does not contain private chain-of-thought, and it is not intended as a ready-made training set. Some configurations contain expected labels or answer mappings because those were part of the registered experimental design. Users should preserve the stated evidence status and split when reproducing or extending the study.

Configuration guide

Readability and open-vocabulary interpretation

Configuration	Rows	Purpose	Evidence status
readability_heldout	50	Ten materials-mechanism families, each expressed in five different phrasings, used for controlled and open-vocabulary hidden-state readout	Frozen held-out suite
readability_development_122	122	Earlier heterogeneous readout suite retained to document hypothesis and method development	Development only; not population evidence
option_free_questions	72	Natural questions without visible answer options, used for option-free state and graph analyses	Frozen existing cohort
automated_interpreter	5	Exact blinded secondary-rater prompts used to classify discovered word sets	Retrospective secondary analysis
lexical_adversarial_discovery	72	Matched prompts designed to separate lexical surface cues from physical relations	Frozen discovery cohort
lexical_adversarial_replication	72	Disjoint replication of the lexical-adversarial design	Frozen replication
answer_code_binding	72	Arbitrary A/B answer-code control used as a prospective falsification test	Failed manipulation check, reported as such

Causal steering

Configuration	Rows	Purpose	Evidence status
steering_initial_pilot	15	Initial A/B intervention pilot retained to document the development history	Development only; not primary evidence
steering_layer_selection	18	Disjoint calibration prompts used in the preliminary candidate-layer comparison	Development only; used for layer selection
steering_preliminary_confirmation	24	Earlier A/B confirmation prompts retained for completeness	Development only; not primary evidence
steering_direction_fit	9	Three prompts per mechanism family used only to fit intervention directions	Frozen direction-fitting set
steering_broad_screen	60	Thirty physical conditions in both answer orders, used for the broad intervention screen	Frozen confirmatory screen
steering_grain_confirmation	24	Prospective grain-refinement/coarsening receivers in both answer orders	Frozen prospective confirmation

Relational physics

Configuration	Rows	Purpose	Evidence status
relational_development	128	Initial 16-law cohort used to develop the relational-state analysis	Development only
relational_fresh_screen	256	Fresh 16-law screen after the development analysis	Frozen fresh screen
relational_disjoint_confirmation	192	Disjoint 12-law confirmation used before freezing the final protocol	Frozen confirmation
relational_final_60	960	Sixty constitutive laws with neutral, direct, and inverse conditions and matched answer-order controls	Frozen final benchmark

Exploratory boundary and composition studies

Configuration	Rows	Purpose	Evidence status
symbolic_monotonicity_pilot	128	Symbolic monotonicity pilot	Exploratory development
symbolic_equivalence_pilot	256	Symbolic-equivalence pilot	Exploratory development
abstract_physics_composition	256	Abstract-code composition follow-up	Exploratory follow-up
abstract_physics_semantic	256	Semantic composition follow-up	Exploratory follow-up

The 21 configurations contain 3,047 rows in total. This number counts prompt appearances across configurations; it should not be interpreted as 3,047 statistically independent scientific concepts.

Shared row schema

All configurations use the same fields:

• system_prompt, user_prompt: exact model-facing text. Most scientific prompts have no separate system prompt.
• expected_answer, answer_options: registered target and visible choices when applicable; empty for genuinely option-free readout prompts.
• family_id, family_name, domain, material_system, case_id, variant: scientific grouping and prompt-construction metadata.
• relation, numeric_direction, law_id, law_sign, answer_order: relational-physics or answer-order factors when applicable.
• prompt_role, evidence_status, split: the role of a prompt in the experiment and whether it was developmental, frozen, confirmatory, or exploratory.
• source_path, source_prompt_id: stable link back to the repository artifact from which the row was exported.
• metadata_json: a deterministic JSON serialization of the full prompt-construction record. Model-generated rater responses are deliberately excluded.

Empty fields mean that the concept does not apply to that configuration; they do not imply missing source data.

Loading the data

Install the Hugging Face datasets library and load one configuration at a time:

python -m pip install datasets

from datasets import load_dataset

heldout = load_dataset(
    "lamm-mit/gemma4-materials-mechanism-prompts",
    "readability_heldout",
    split="test",
)

final_laws = load_dataset(
    "lamm-mit/gemma4-materials-mechanism-prompts",
    "relational_final_60",
    split="test",
)

print(heldout[0]["user_prompt"])
print(final_laws[0]["law_id"], final_laws[0]["expected_answer"])

To inspect every configuration programmatically:

from datasets import get_dataset_config_names

print(get_dataset_config_names("lamm-mit/gemma4-materials-mechanism-prompts"))

The configuration selector in the Hugging Face dataset viewer is the simplest way to browse the sections interactively.

Rebuilding and validating the release

From the root of the accompanying GitHub repository:

python scripts/build_hf_prompt_dataset.py
python scripts/validate_hf_prompt_dataset.py

manifest.json records the row count and SHA-256 digest of every exported JSONL file, together with the source file and its digest. The hashes are provenance checks for this release; they are not scientific outcomes.

Derived analyses such as graph topology, graph isomorphism, community structure, GIN tests, and activation patching reuse the registered prompt cohorts rather than creating undocumented prompt sets. prompt_reuse_map.json records those links without duplicating the prompt rows.

Related artifacts

• Jacobian-lens checkpoints and run artifacts: lamm-mit/gemma4-jacobian-lenses
• Held-out latent vectors: lamm-mit/gemma4-materials-latent-vectors
• Source code, frozen protocols, results, and supplementary artifacts: lamm-mit/Substrates

Scope and responsible use

These prompts were designed to study one open-weight model and a specific materials-science interpretability workflow. They are not a comprehensive materials benchmark, and the expected labels should not be treated as expert-reviewed ground truth beyond the scope described in the paper. If the corpus is used for training or prompt selection, evaluate on a genuinely disjoint cohort and report that reuse, since contamination would invalidate the held-out and confirmatory roles documented here.

Citation

The manuscript is being prepared for submission. Please cite the paper and this dataset release once the archival identifiers are available.

License

This dataset is released under the Creative Commons Attribution 4.0 International license.

Citation

@misc{buehler2026readingsteeringrepresentationsmaterialsscience,
      title={Reading and Steering Representations of Materials-Science Mechanisms in an Open-Weight Language Model}, 
      author={Markus J. Buehler},
      year={2026},
      eprint={2607.20058},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2607.20058},
}