func_localize_claude45_1457i_text50

Verbosity-ablation variant of synthetic-code-training/func_localize_claude45_1457i: the prose before every tool call is rewritten by nvidia/deepseek-ai/deepseek-v4-flash to about 50 tokens (accepted band 38-62 tokens of the Qwen3 tokenizer, up to 3 rounds; 8 of 26194 turns missed the band and keep their original prose).

Construction (shared by all _text* siblings): think and task_tracker turns and their result turns are removed (trajectories contain only real tool calls); the system prompt, task, tool calls and tool results are byte-identical to the base. The rephraser saw only the current turn (its prose + its tool call); the ~300-token version was written first and condensed to 100/50/20 in the same response so the four lengths share one meaning. Turns with no original prose received prose explaining their tool call. Malformed tool calls (garbled <tool_call> JSON / <invoke>) are kept verbatim as the call part.

	value
rows	1457
assistant turns (excluding think/task_tracker)	26194
think/task_tracker turns removed	4421
prose tokens per turn: mean / median	50.1 / 50
turns with empty prose	1
turns kept original prose (band missed)	8

Built with tools/verbosity_rephrase in the benchmarks repo.