Qwen2.5-7B-Instruct Trips on Summaries

This looks less like a temperature problem and more like a grounding problem: a 7B instruct model can rewrite text, but semi-structured summarization needs strict constraints or it will fabricate connective tissue.

• –Qwen2.5-7B-Instruct has a huge advertised context window, so the bottleneck is not raw token capacity
• –Community replies favor smaller newer models like Gemma and Qwen3.5, plus IBM Granite, which has a better reputation for summarization behavior
• –The robust pattern here is hierarchical summarization: extract facts per note or tag first, then synthesize the final report in a second pass
• –Add a hard no-new-facts rule, require tag-by-tag coverage, and make the model cite or paraphrase only the source notes
• –For this use case, evaluate hallucinated entities and missed themes, not just fluency or compression ratio