GREP · Software mentions — V1 (GPU)

Reads scientific PDFs and pulls out every software name, tool and library the authors used, with its location in the text, a type and a purpose. This leaf runs V1, the crowd's first-generation model. GREP runs three independently trained models over the same papers; a mention enters the final record only when at least two of the three find the same span. The GPU build produces exactly the same mentions as the CPU build and clears a unit in two to three minutes instead of twenty. A work unit is about ten papers, roughly 5 MB in and one small JSON out. Every unit goes to three different volunteers and the head accepts it once two produce matching output. Needs an NVIDIA GPU with 4 GB of VRAM, plus 6 GB of free memory and 20 GB of free disk. The model ships inside a one-time 5.6 GB image download, after which the work runs entirely offline.

Measured accuracy (character-level F1; higher is better):

Model dev test Softcite SoFAIR clean

f14 0.823 0.819 0.8653 0.8909 f13 0.814 0.807 0.8637 0.8854 V1 0.800 0.781 0.8096 0.7950 <- this leaf crowd (2-of-3) 0.833 0.828 0.8648 0.8940