danielrosehill/term-context-sentences

Term-Context Sentences

Training data for danielrosehill/term-context-sentence-generator-granite-4.2-3b-v1: given a term, a one-line context and a count n, write n natural English sentences that each use the term once, in the meaning the context describes. The sentences are short (6 to 18 words) and are made to be read aloud when recording speech-recognition (ASR) training data. The terms are romanized Hebrew and Yiddish words, Israeli place, street and brand names, people's names, and tech and AI vocabulary.

Configs

config	rows	row
train	13,396	term, context, kind (optional), n, sentences (exactly n), source (seed / topup), messages (the chat-format SFT example)
eval	307	term-disjoint from train: term, context, kind, references (sentences, may be empty), bench
contexts	8,112	term, context, kind, confident, source (the model that wrote it)

kind values: hebrew, yiddish, name, place, street, brand, drug, medical, tech, ai_model, linux, smart_home, jewish_text, english_word, other.

How it was made (2026-11-03)

• Sentences: 26,331 existing single-term sentences from two ASR projects (an English-Hebrew code-switching sentence bank and a voice-capture app's task prompts). Most were written by LLMs (Claude Sonnet, Haiku and Opus, DeepSeek, Gemini) with some handwritten. Added to that: 12,202 sentences for 800 terms from Claude Sonnet 5.5, so the set includes large-n examples.
• Filter: every sentence passes an ASR acceptance gate. That means 6 to 18 words, the term exactly once (word boundary), a restricted ASCII charset, no digits outside the term, no number of 10 or more written in words, and code-switching and tech terms never mixed in one sentence. About 22% of the seed sentences failed.
• Contexts: 8,116 terms had no context. Gemini 3.1 Flash-Lite wrote one from each term and its own sentences, so the context matches the sense the sentences use. It was picked over 11 other models in a blind-judged comparison. The other terms use the definitions from the source vocabulary packs.
• Examples: each training term gives one or more examples, with different subsets, sizes (n = 1 to 20) and orders of its sentences.
• Removed for publication: rows naming the author, his family or his company.

Caveats

Synthetic, single-speaker oriented (an English speaker in Jerusalem), and Israel-specific in vocabulary. Contexts are model-written and not hand-checked. Sentences pass the gate, but the gate checks form, not meaning.