ECLIPTICA-30k

ECLIPTICA holds the user prompt fixed and varies only the alignment instruction, so a systematic behavioral difference between rows with the same prompt_id is attributable to the instruction. ECLIPTICA-30k crosses the same 300 prompts as ECLIPTICA-3k with 100 instruction types (10 axes x 10 settings) = 30,000 cases.

Provenance. The 300 prompts are author-written policy and ethics questions (12 categories x 25), and all 100 instruction types, the original 10 included, were written by the authors. No judge models, annotators or external preference data were used to create prompts or types.

What has been evaluated. Every ECLIPTICA result in the companion paper uses the original 10 types, i.e. the eval_3k config (300 x 10 = 3,000 cases, byte-identical to ECLIPTICA-3k). The 90 additional types are released to widen the instruction space for further research; like the original 10 they are hand-designed, they have not been scored by any model yet, and they carry no evaluation claims.

Corrections (2026-10-20). Earlier versions of the ECLIPTICA-3k card said the prompts came from HH-RLHF and the 10 original types were derived by judge models, and this card's wording implied the same. Both were wrong: prompts and types are author-written (src/eval/isd_dataset.py in the code repository). The data files are unchanged.

Configs

config	rows	content
full (default)	30,000	300 prompts x 100 types; adds axis, axis_kind, origin columns
eval_3k	3,000	the evaluated subset, identical to ECLIPTICA-3k (6 columns)

Axes and settings

kind = policy axes change what the model commits to (stance, boundary, epistemics, role, region); kind = surface axes change how it says it (style, length, audience, format). Each type has three imperative phrasings; one is sampled per row with a fixed seed.

axis	kind	settings	types
stance	policy	10	neutral, conservative, liberal, utilitarian, deontological, libertarian, communitarian, precautionary, technocratic, egalitarian
constraint_posture	policy	10	regulatory_aware, safety_first, privacy_first, harm_reduction, child_safe, legal_conservative, permissive_informational, medical_caution, financial_caution, zero_speculation
delivery_style	surface	10	empathetic, educational, concise, professional, creative, casual, humorous, socratic, narrative, diplomatic
epistemic_stance	policy	10	hedged_uncertainty, confident_assertion, evidence_cited, consensus_only, skeptical, speculative_labeled, quantitative, probabilistic_calibrated, devils_advocate, first_principles
verbosity_structure	surface	10	one_sentence, three_bullets, numbered_steps, executive_summary, exhaustive, tldr_first, faq, outline_headings, comparison_table, hundred_words
audience	surface	10	young_child, teenager, undergraduate, domain_expert, policymaker, executive, older_adult_nontechnical, plain_english_esl, journalist, software_developer
persona_role	policy	10	support_agent, legal_advisor, clinician, teacher, financial_advisor, compliance_officer, research_scientist, product_manager, community_moderator, counselor
action_orientation	policy	10	recommend_one, options_only, risk_assessment, clarify_first, step_by_step_plan, tradeoff_matrix, stakeholder_map, historical_context, future_scenarios, counterargument_first
regional_lens	policy	10	global_south, india, european_union, united_states, east_asia, rural, urban, small_business, public_sector, nonprofit
format_constraints	surface	10	markdown_headings, json_output, plain_text, inline_citations, glossary, worked_examples, counterexamples, closing_summary, definition_first, quantitative_estimates

The full taxonomy with descriptions, phrasings and expected characteristics is in taxonomy.yaml.

Fields

• prompt_id: identifier shared by the 100 rows of one prompt
• prompt: the user query (held fixed across instruction types)
• instruction_type: one of the 100 types above
• instruction: the alignment-instruction text used for this row
• expected_characteristics: behavioral traits the response should show under this instruction
• source: topical category of the prompt (12 categories x 25 prompts)
• axis, axis_kind, origin (full config only): taxonomy axis, policy/surface, and whether the type belongs to the evaluated 10

Usage

from datasets import load_dataset
full = load_dataset("anonymousML123/ECLIPTICA-30k", "full", split="train")       # 30,000 rows
eval_3k = load_dataset("anonymousML123/ECLIPTICA-30k", "eval_3k", split="train") # the evaluated 3,000 rows

Citation

@misc{wanaskar2026ecliptica,
  title         = {ECLIPTICA: A Framework for Switchable LLM Alignment via SwiPO, Switchable Preference Optimization},
  author        = {Wanaskar, Kapil and Jena, Gaytri and Jain, Vinija and Chadha, Aman and Das, Amitava},
  year          = {2026},
  eprint        = {2601.06157},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2601.06157}
}

License

CC-BY-4.0.