Views
No views yet
script.py; the only change is the model. Six submissions built on
Qwen2.5-14B-Instruct-AWQ showed the score was bounded by the model failing to
solve the problems (exact match ~0.025 against the organizers' 0.0729 with the
same weights), not by parsing or prompting — seven parser variants replayed over
identical raw generations all scored exact match 0.0000.microsoft/phi-4 is loadable under the sandbox's pinned transformers 4.44.1
(Phi3ForCausalLM, rope_scaling: None, standard AWQ gemm, no remote code),
fits the T4 at 9.1 GB, and unlike Qwen2.5 ships no repetition_penalty in its
generation config.stelterlab/phi-4-AWQ.