Reference
Prompts
Every SenseBench run uses one of these registered prompt formats to query the model. A prompt fixes the context window, candidate-sense ordering, target marker, and output format, so runs on the same prompt are directly comparable. Open a prompt for full metadata, the message templates, rendered item examples, and a download of the raw JSON.
| ID | Name | Description |
|---|---|---|
| p001 | 5+1 Context, Frequency-Ordered WordNet Glosses | English fine-grained Word Sense Disambiguation prompt using the target sentence with 5 preceding and 1 following sentence of context. The context is detokenized from the Penn-Treebank-style source tokens into natural English (punctuation, contractions, and quotation marks restored) before the target is marked with <t>...</t>. Candidate senses include WordNet sense keys, definitions, synonyms, and usage examples, ordered by WordNet frequency; the model returns JSON with sense_index. |
| p002 | Minimal Single-Sentence Prompt | Minimal English Word Sense Disambiguation prompt: only the target sentence, detokenized from the Penn-Treebank-style source tokens into natural English (punctuation, contractions, and quotation marks restored) before the target is marked with <t>...</t>. Candidate senses give an index, definition, and at most one usage example, ordered by WordNet frequency; the model returns a plain integer. |
| p003 | p002 + 5+1 Context + Synonyms + Two Examples | Workhorse English Word Sense Disambiguation prompt selected from the p021 ablation candidate: based on p002 with the p001 context window, synonyms, and a second candidate example per sense added while keeping no target metadata, no WordNet sense keys, and plain-number output. |
| p004 | 5+1 Context, Randomized WordNet Glosses | English fine-grained Word Sense Disambiguation prompt matching p001 except candidate senses are deterministically randomized per item instead of ordered by WordNet frequency. The target sentence uses 5 preceding and 1 following sentence of detokenized context, the target is marked with <t>...</t>, candidate senses include WordNet sense keys, definitions, synonyms, and usage examples, and the model returns JSON with sense_index. |