Raw code generations from the untrained base model (Qwen2.5-Coder-14B-Instruct)
evaluated on the Synergistic Dependency Selection (SDS) benchmark.
Contents
generations.jsonl: 64,000 lines (64 samples x 1,000 problems)
Seed: 101
Model: Qwen/Qwen2.5-Coder-14B-Instruct (no RL fine-tuning)
Decoding: temperature=0.6, 64 independent samples per problem
System prompt: Identical to Hero (from config_hero.yaml)