Atlas3D/character-creation-big-five

Character Creation — synthetic Big Five scenarios

40,072 original model-generated prompts for exploring LLM behavior. Published by Atlas3D / Character Creation. This is a growing discovery dataset; the Hub commit pins each snapshot.

What the labels mean

trait_design and facet_design describe the generation assignment, not a measured personality score, answer key, validated inventory, or human annotation. The six facets per trait are our operational prompt taxonomy. All records are discovery material, exposed as train for dataset-loader compatibility. No held-out evaluation split is claimed; do not use these records both to tune and to claim generalization.

Trait counts: {"agreeableness": 9372, "conscientiousness": 7895, "extraversion": 6920, "neuroticism": 7250, "openness": 8635}.

Generation and filtering

A local 27B planner assigns trait, facet and domain; six Qwen3.8-27B-NVFP4 workers produce batches of twelve scenario prompts. Generator checkpoint: nvidia/Qwen3.8-27B-NVFP4, revision 482ca0f3832238542f8f5295dde86b5f22711d80 (architecture qwen3_5). Temperature 0.9, top-p 0.95, top-k 40, maximum 2048 output tokens per batch, thinking disabled. Seeds vary by batch and worker. Request and response hashes, job IDs, and exact model/runtime identities are retained in each row; the operator retains full request audits. Validation checks structure and prompt constraints; normalized-text deduplication rejects exact normalized duplicates. Semantic near-duplicates, bias, uneven trait coverage, and weak scenarios can remain. There has been no independent psychometric or human-quality validation.

Fields and reproducibility

prompt is the text to present to an LLM; trait_design, facet_design, and domain are design metadata. id and hashes support stable joins and provenance. synthetic is always true and validated_inventory always false. manifest.json records source/export SHA-256 and row counts. No local filesystem paths, credentials, book text, SCOTUS documents, or questionnaire item collections are included in this export.

from datasets import load_dataset
scenarios = load_dataset("Atlas3D/character-creation-big-five", split="train")

Attribution and use

The neuron-steering research workflow is inspired by Lyra / neuron-explorer, with runtime inspiration from RhizoNymph/vLLM and heterodoxin/Strata. These prompts were generated by the Atlas workflow; Lyra did not author or validate them. Credit Atlas3D / Character Creation and link this dataset when redistributing under CC BY 4.0. This dataset is intended for LLM research, not psychological diagnosis or consequential assessment of people.

Research captures and evidence

The original scenarios.jsonl, its manifest.json, and default dataset configuration remain unchanged. See human guide and agent guide.

• research_assets.json indexes the shared model configuration, authored probes, public presets and calibrated layer arrays with hashes.
• research_manifest.json indexes immutable JSONL batches, record IDs, SHA-256 hashes and exact publication coverage.
• Activation rows include public prompt alignment, encoded tokens and complete top-32 per-token profiles. They are not full all-neuron arrays.
• research_raw_manifest.json separately indexes automatically uploaded immutable tar shards of original validated NPZ arrays and entry manifests. Most recently observed verified raw coverage: 41,208 records; consult that manifest for current coverage. Pending arrays must not be treated as uploaded.
• Raw prefill reductions cover every neuron and encoded prompt position, including template tokens; they are reductions, not a full neuron-by-token tensor. Baseline answer captures retain final-prefill arrays, not answer-token activation trajectories.
• Only the exact approved synthetic prompts and authored probes are exported. Book material and arbitrary user prompts are excluded; other tiers appear only as aggregate counts.
• Associations and calibrated probe precision are not causal neuron meanings or validated human personality scores.