ProCreations/chilltext-v2

ChillText v2

ChillText v2 is a 200,000-example synthetic style-transfer dataset for making an assistant's existing answer sound like a real, current Discord or iMessage reply without changing its meaning. It adds the original user message as context, optional natural-language personality control, strict preservation of links and identifiers, and four times as many rows as ChillText v1.

Format

{"system":"","input":"the user's actual question","agentOutput":"the assistant's normal answer","output":"casual message 1\n\ncasual message 2","n_msgs":2,"batch_id":123}

• system: optional English personality request. An empty string selects the original ChillText personality: lowercase, extremely casual, no emojis or em dashes, and restrained, context-appropriate current slang.
• input: the user's real message or question to the assistant.
• agentOutput: the assistant's normal answer before style transfer.
• output: the same answer rewritten in ChillText style. Separate chat bubbles are joined by a blank line (\n\n).
• n_msgs: number of output bubbles, always 1–4.
• batch_id: generation bookkeeping.

The optional system prompt changes personality rather than factual content. Examples include warm, dry, snarky, concise, patient, deadpan, playful, or direct styles. A deliberately small subset explicitly permits ordinary profanity. Those prompts teach occasional, context-aware swearing while prohibiting slurs, hate, threats, and abusive targeting. Profanity is rejected in all empty-system and ordinary-system rows.

Release statistics

measure	value
rows	200,000
default/empty-system rows	121,551
custom-system rows	78,449
profanity-permitted system rows	4,061
rows containing exact-copy invariants	80,216
1 / 2 / 3 / 4 message rows	68,635 / 98,734 / 26,566 / 6,065

Quality controls

Rows were generated with GPT-5.6 Luna at low reasoning. Reasoning was ignored at the transport layer and is not included in the dataset. Every retained row passed deterministic checks for:

• exact preservation of URLs, emails, paths, commands, model names, versions, dates, times, prices, percentages, measurements, ticket/order IDs, and other numeric identifiers;
• coherent 1–4-message output formatting;
• no emoji, em dash, en dash, or dated/forced internet slang;
• lowercase default output except case-sensitive exact-copy spans;
• no duplicate interactions and no unchanged agentOutput/output pairs;
• profanity only when the explicit system prompt permits it.

Release candidates also received a Luna major-defect review for incoherence, contradictions, invented or unrelated details, changed uncertainty or refusal behavior, and safety-critical omissions. Introduced-anchor candidates received a separate relevance review. Rows rejected by either review were regenerated, and every replacement tail was reviewed again before release.

Automated validators cannot prove perfect semantic equivalence. Applications should keep their normal safety checks and evaluate the model on their own domain, especially for medical, legal, financial, or other high-stakes text.

Suggested training prompt

Use system as the model's system message when it is nonempty, omit the system turn otherwise, and send this user message:

rewrite the agent output in chilltext style using the user's message as context.

user:
{input}

agent output:
{agentOutput}

return only the rewritten chat message or messages, separated by a blank line. use 1 to 4 messages total and never more than 4.

Train only on output as the assistant completion. Gemma 4 natively supports system messages, so personality prompts can be supplied directly in the system role.

Generation and intended use

The data is intended for supervised style-transfer training and evaluation. Facts are synthetic; do not treat examples as real news, medical records, transactions, or operational instructions. The .example.com/.example resources and 555 phone numbers are deliberately non-production.