The model wrote six sentences about a grandmother. Two were allowed on the altar.

Описание и применение

Инструмент генерирует памятные предложения, проверяя каждое на наличие корректных цитат и фактической поддержки, чтобы избежать вымышленных деталей.

Задача

A language model can hallucinate personal details when generating memorial text, risking false memories in a sensitive context.

Результат

The tool produces a single HTML altar file with six sentences, of which two are fully verified and the rest are either repaired or omitted, eliminating hallucinated personal details.

Рабочий процесс

• Pull local model with ollama pull llama3.2:1b
• Run python3 ofrenda.py --memories examples/memories.json --out altar.html
• Model generates sentences with numbered citations
• Python script validates each sentence: checks citations exist, matches numbers, names, and content words
• If a sentence fails, it is repaired using only cited facts or dropped
• Output is a single self‑contained HTML file with verified sentences

Доказательства

One command, one self‑contained HTML file, no API key, no account, no network.

Источник

The rule: a sentence must cite the memory it came from

Источник

Anything that fails gets one repair attempt against only the facts it cited, and then it is dropped.

Источник