{
  "date": "2026-09-11",
  "task": "English sentence-completion grammar; not factual correctness or Polish quality",
  "raw_reference": "../english-grammar-100-2026-09-11/",
  "raw_preserved": true,
  "chat_samples": 100,
  "checkpoint_sha256": "c16841621efe93c7c8ebf1b374709a96276f3741e649f83ed90131a7b5ad23a8",
  "chat_template": "{{bos_token}}{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
  "user_instruction": "Complete the English sentence started by the assistant. Continue directly from the supplied prefix without explanations.",
  "assistant_prefill": "Exact original English prefix appended after checkpoint assistant generation marker",
  "sampling": {
    "temperature": 0.8,
    "top_k": 40,
    "top_p": 1,
    "min_p": 0,
    "max_new_tokens": 160,
    "cache_prompt": true
  },
  "judge": "Qwen3.8-27B W4A16; original large_judge.py rubric; temperature 0; model/configuration hidden",
  "judge_system_sha256": "a82d293bd534cf6efc0a4af916f0980cf97e6c0515f96256736efdd2fe67a995",
  "limitations": [
    "Compares raw completion with instruction plus chat template plus assistant prefill, not template delimiters alone.",
    "Single seed per prefix; raw reference is an earlier frozen run.",
    "Automatic grammar/completeness judgments are not human validation and do not assess factual correctness.",
    "These are English tasks; Polish-trained models in the scatter are English transfer probes.",
    "Wilson intervals exclude judge error and sampler nondeterminism."
  ]
}