func_localize_claude45_1457i_text32x

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 32 times its own length in Qwen3 tokens (accepted band ±25 %, written in consecutive parts of at most 1500 tokens, each request continuing the text so far, starting from the accepted 8x version as its source). Of 26194 turns, 15754 were not rephrased (empty prose, or a target outside 4-8000 tokens) and 49 missed the band; both keep their original prose.

Construction (shared by the _text0x/0.5x/2x/4x/8x/32x siblings): think and task_tracker turns are kept verbatim, as are the system prompt, task, tool calls and tool results. For the rephrased siblings the rephraser saw only the current turn (its prose + its tool call) and wrote the four versions in one response as a ladder (0.5x shortens the original, 2x elaborates it, 4x elaborates the 2x, 8x elaborates the 4x), so the four lengths share one meaning; the 32x version was written afterwards by elaborating the 8x version further, in parts. A turn's targets are multiples of its own prose length, so empty turns stay empty. 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 kept verbatim	4421
prose tokens per turn: mean / median	467.7 / 0
prose tokens / base prose tokens, rephrased turns: mean / median	33.46 / 32.70
turns not rephrased (empty prose or target outside 4-8000 tokens)	15754
versions whose verbatim-repeated sentences were removed afterwards (dedupe_parts.py) / of them left under the band	544 / 12
turns with empty prose	15681
turns kept original prose (band missed)	49

Built with tools/verbosity_rephrase in the benchmarks repo.