Data: BEGIN , by Nouha Dziri, Hannah Rashkin, Tal Linzen, David Reitter, CC BY 4.0 . The replies of the benchmark's Wizard of Wikipedia part, from conversations by Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli and Jason Weston at Facebook, without those labelled Generic: its development set, and 1,000 of its test set drawn at random, each with its set, the model that wrote it, its label, its sentence as `context` and itself as `reply`.
Data: BEGIN , by Nouha Dziri, Hannah Rashkin, Tal Linzen, David Reitter, CC BY 4.0 . For each reply in begin.csv, the probabilities Granite Guardian 3.1 2B at 8 bits gave the answers Yes and No to whether it keeps to its sentence, summed over its 20 likeliest first tokens, through llama.cpp's server, release b11734, on 20 October 2026.
Data: MultiDoc2Dial , by Song Feng, Siva Sankalp Patel, Hui Wan, Sachindra Joshi, CC BY 3.0 . The dataset's 488 help pages in passages: each page's text taken from its HTML, with its headings marked by #s and its list items by -, and its blocks joined into passages of up to 200 words. Each passage has its page, the page's domain, and the page's title without the part number the dataset adds.
Data: MultiDoc2Dial , by Song Feng, Siva Sankalp Patel, Hui Wan, Sachindra Joshi, CC BY 3.0 . The opening question of each of the dataset's test dialogues whose next turn answers it from the same page: 340 questions, each with its page and the page's domain.
Data: MultiDoc2Dial , by Song Feng, Siva Sankalp Patel, Hui Wan, Sachindra Joshi, CC BY 3.0 . For each question, the passages that hold at least half of a part of the page its answer was given from, and every passage with the same text as one of them.
Data: MultiDoc2Dial , by Song Feng, Siva Sankalp Patel, Hui Wan, Sachindra Joshi, CC BY 3.0 . For each question, the first 10 passages of the help desk's reranked search: the first 20 texts of a reciprocal rank fusion, with k = 60, of the passages' BM25 ranking and their vectors' ranking, in the order of the scores Qwen3-Reranker-0.6B at 8 bits gave them, through llama.cpp's server, release b11734, on 20 October 2026.
Data: MultiDoc2Dial , by Song Feng, Siva Sankalp Patel, Hui Wan, Sachindra Joshi, CC BY 3.0 . For each question, the reply the dataset's agent gave it in the next turn.
Data: MultiDoc2Dial , by Song Feng, Siva Sankalp Patel, Hui Wan, Sachindra Joshi, CC BY 3.0 . For each reply in human-replies.csv, the probabilities Granite Guardian 3.1 2B at 8 bits gave the answers Yes and No to whether it keeps to its context, summed over its 20 likeliest first tokens, through llama.cpp's server, release b11734, on 21 October 2026: with the passages of its question's page that hold the answer; with the first passage of the reranked search on another page that holds none of their texts; and with the passages that hold the answer to another service's question.
Data: MultiDoc2Dial , by Song Feng, Siva Sankalp Patel, Hui Wan, Sachindra Joshi, CC BY 3.0 . For each question, the reply Qwen3.5-2B at 8 bits wrote from its 5 passages, through llama.cpp's server, release b11734, on 21 October 2026, at the sampling settings its model card gives for text, with a seed of its own.
Data: MultiDoc2Dial , by Song Feng, Siva Sankalp Patel, Hui Wan, Sachindra Joshi, CC BY 3.0 . For each reply in replies.csv, the probabilities Granite Guardian 3.1 2B at 8 bits gave the answers Yes and No to whether it keeps to the passages it was written from, summed over its 20 likeliest first tokens, through llama.cpp's server, release b11734, on 21 October 2026.
Model: Granite Guardian 3.1 2B's tokenizer configuration , by IBM, Apache License 2.0 .
Model: Granite-Guardian-3.1-2B , by IBM, Apache License 2.0 .