Dataset Viewer
Auto-converted to Parquet Duplicate
instance_id
stringlengths
15
51
resolved
bool
1 class
messages
listlengths
16
122
Gallopsled__pwntools-636_92
true
[{"content":"You are OpenHands agent, a helpful AI assistant that can interact with a computer to so(...TRUNCATED)
ipython__ipython-374_620
true
[{"content":"You are OpenHands agent, a helpful AI assistant that can interact with a computer to so(...TRUNCATED)
Gallopsled__pwntools-636_94
true
[{"content":"You are OpenHands agent, a helpful AI assistant that can interact with a computer to so(...TRUNCATED)
Gallopsled__pwntools-636_206
true
[{"content":"You are OpenHands agent, a helpful AI assistant that can interact with a computer to so(...TRUNCATED)
Gallopsled__pwntools-636_95
true
[{"content":"You are OpenHands agent, a helpful AI assistant that can interact with a computer to so(...TRUNCATED)
ipython__ipython-374_821
true
[{"content":"You are OpenHands agent, a helpful AI assistant that can interact with a computer to so(...TRUNCATED)
ipython__ipython-374_309
true
[{"content":"You are OpenHands agent, a helpful AI assistant that can interact with a computer to so(...TRUNCATED)
ipython__ipython-374_438
true
[{"content":"You are OpenHands agent, a helpful AI assistant that can interact with a computer to so(...TRUNCATED)
Gallopsled__pwntools-636_316
true
[{"content":"You are OpenHands agent, a helpful AI assistant that can interact with a computer to so(...TRUNCATED)
ipython__ipython-374_456
true
[{"content":"You are OpenHands agent, a helpful AI assistant that can interact with a computer to so(...TRUNCATED)
End of preview. Expand in Data Studio

func_localize_claude45_1457i_text32x_think32x

Verbosity-ablation variant of synthetic-code-training/func_localize_claude45_1457i: the prose before every tool call and the thought of every think call are rewritten by nvidia/deepseek-ai/deepseek-v4-flash to 32 times their own length in Qwen3 tokens (accepted band ±25 %); task_tracker turns are kept verbatim. Prose: of 29932 turns, 18975 were not rephrased (empty, or a target outside 4-8000 tokens) and 53 missed the band. Thoughts: of 3738, 0 were not rephrased (a target outside 4-16000 tokens) and 277 missed the band. All of them keep their original wording. A thought whose 32x would exceed 16000 tokens is rewritten to 16000 tokens instead (longer versions degenerate into filler).

Construction (shared by the _text<N>x_think<N>x siblings): the same as the _text0x/0.5x/2x/4x/8x/32x siblings - the prose of every non-think turn is the very same rewritten text - except for the think turns: their prose and the thought argument of the call are rewritten to the same multiple (_text0x_think0x removes them), while task_tracker turns stay verbatim. The system prompt, task, tool results and every other tool-call argument are byte-identical to the base. The thought rephraser saw only the thought and the call's summary. 0.5x and the versions of up to 600 tokens were written in one response; every longer version was written in consecutive parts of at most 1500 tokens, elaborating the version below it (4x from 2x, 8x from 4x, 32x from 8x). A text's targets are multiples of its own length, so empty prose stays empty. Malformed tool calls (garbled <tool_call> JSON / <invoke>) are kept verbatim as the call part.

value
rows 1457
assistant turns (excluding task_tracker) 29932
think turns removed / task_tracker turns kept verbatim 0 / 683
prose tokens per turn: mean / median 431.7 / 0
prose tokens / base prose tokens, rephrased turns: mean / median 33.48 / 32.71
thoughts: rephrased / not rephrased / kept original (band missed) 3461 / 0 / 277
thought tokens: mean / median 8556.0 / 8892
thought tokens / base thought tokens, rephrased thoughts: mean / median 30.86 / 30.84
versions whose verbatim-repeated sentences were removed afterwards (dedupe_parts.py) / of them left under the band 544 / 12
turns with empty prose 18895
turns kept original prose (band missed) 53

Built with tools/verbosity_rephrase in the benchmarks repo.

Downloads last month
32