Alberto1231/prism_trial_3

PRISM Trial 3

A simplified next-human-message benchmark derived from the conversations configuration of HannahRoseKirk/prism-alignment.

Each example contains a genuine human question, the model response selected by that participant (if_chosen == true), and the genuine human message from the next interaction turn. Rejected model alternatives are excluded.

Splits

• train: 5 examples intended only for few-shot prompting
• test: 456 scored examples

All 461 rows have distinct user_id and conversation_id values. The test split is exactly balanced across PRISM's three conversation types: 152 controversy guided, 152 unguided, 152 values guided.

The scored split is a near-census of the eligible source pool: 456 of 461 eligible surveyed respondents are in test, and the other 5 supply the fixed demonstrations. Demographic levels are not hard quotas, so a rare level cannot shrink the benchmark or cause duplicate-person oversampling. Downstream subgroup comparisons use equal-size, fixed-seed respondent subsets separately within each sufficiently supported attribute. Exact joint demographic balance is neither claimed nor required.

The nine documented self-reported demographic fields are published with each row for transparent subgroup analysis; they are not inferred and are not included in model prompts.

Prompt and target

Conversation:
Human: {human_question}

Assistant: {llm_response}

Human: 

The target contains only the next genuine human message and is never included in the prompt.

Filtering and sampling

• Source revision: 18ab5cfb37456f4ec8cbc00212ce54cf7b1239f6
• Random seed: 42
• Source conversation date: 2023 or later
• Target length: 30 to 200 Unicode words
• Every target fits the 256-token generation cap with the pinned tokenizer
• Individual formatted context: at most 400 tokens with the pinned MARIN tokenizer
• Complete five-shot input: at most 1,792 tokens, reserving 256 tokens for generation in models with a 2,048-token context window
• English and PII filtering: PRISM's source-provided per-text metadata
• At most one example per user and conversation across both splits
• Random selection after predetermined filters operates on shuffled unique users rather than transitions, so prolific participants do not receive extra sampling weight
• Exact balance across the three conversation types in the scored split
• Rare demographic levels remain eligible and never impose a global hard quota; exact joint balance is generally infeasible

Fields

prompt, target, date, conversation_id, user_id, conversation_type, source_model, human_question, llm_response, target_word_count, age, gender, study_locale, lm_familiarity, lm_frequency_use, education, employment_status, marital_status, and english_proficiency.

The nine demographic fields are normalized versions of self-reports in the pinned PRISM participant survey. They describe the author of target. No demographic attribute was inferred from text. Rare levels remain in the data and do not impose a global sampling quota.

Licensing

PRISM states that human-written text is licensed under CC BY 4.0 and model responses under CC BY-NC 4.0. Use of model responses must also comply with the original model providers' terms. This derived dataset therefore uses license: other rather than representing the mixed-source material with one license identifier.