GGUF conversion for llama.cpp (Q4_K_M, Q6_K, Q8_0)
Hi, and thanks for releasing WebJev.
I converted it to GGUF so it can run with llama.cpp on smaller machines: https://huggingface.co/emoubarak/WebJev-35B-A3B-GGUF
llama.cpp's /v1/systemone route does not support the decider prompt format yet, so the repo ships your serving adapter adapted to llama-server: same prompt code and answers, only the forward pass changes (a one-token completion whose grammar allows only the option labels).
Quick results on an A100 with jev-ultrafast: each of the three files completed 15 of 15 runs on five practice-site tasks, and Q4_K_M made the same decision as Q8_0 on 52 of 54 recorded browser steps. Latency and JevBench numbers are on the model card.
Feel free to link it from your model card if it helps. Happy to change anything.