TITLE
Local Tab-Title Generation Under Product Constraints: An Eleven-Session Study of Data, Decoding, and Distillation

AUTHORS
MD Ishtiaque Hossain

PRIMARY CATEGORY
cs.CL

SUGGESTED CROSS-LIST
cs.LG

COMMENTS
20 pages, 4 figures, 9 tables. Longitudinal empirical study of local, product-constrained tab-title generation using 12.7M- to 77M-parameter T5-family systems.

ABSTRACT
Generating a useful title for a short interactive session is an extreme summarization problem: the output must identify intent in one to three words, remain grounded in sparse user text, satisfy a 32-character interface contract, and run locally inside a desktop application. We report an eleven-session longitudinal study covering 12.7M-, 24.5M-, 35M-, and 77M-parameter T5-family systems, extractive post-processing, source-constrained beam search, judged supervised fine-tuning, pointer heads, local selectors, and multi-teacher pseudo-labeling. On a sealed Stack Overflow proxy, changing only decoding improved an unchanged 35M model by 0.45-0.48 judge points relative to corrected greedy search and produced a 0.752-point paired advantage over a title-tuned FLAN baseline. However, the same system failed a pre-specified absolute-usefulness audit, and its proxy advantage did not transfer to real terminal tasks. On the terminal domain, a 24.5M encoder-plus-pointer system beat the 35M decoder by 0.13 points at 3.88 ms, while a domain-tuned 77M FLAN model gave the best quality among the measured candidates. Continuing that model on a leak-checked 6,219-example multi-teacher mixture improved two unseen 1,000-task holdouts by 0.21 and 0.20 points, reaching a 7.59 mean score and an 85.7% solid-title rate on the first holdout. The unchanged graph measured 29.3 ms per title in the preceding session. In contrast, 175,642 weak teacher labels and lower validation cross-entropy did not improve product quality. Across this campaign, decoding, target dialect, data curation, board separation, and checkpoint-specific composition mattered more than raw label count or training loss. We preserve negative results and distinguish within-packet evidence from cross-packet observations, offering a case study in product-constrained small-model research rather than a universal benchmark claim.

LICENSE
Choose during arXiv submission.