scripts/results-longrep-prod-topup.json

{
 "model": "glm-5.3-flash",
 "ctx": 300000,
 "gen": 2200,
 "rounds": [
  {
   "round": 1,
   "prompt_tokens": 259632,
   "completion_tokens": 2712,
   "wall_s": 69.2,
   "finish_reason": "length",
   "four_gram_max": 2,
   "four_gram_top": "as a function of",
   "eight_gram_max": 1,
   "eight_gram_top": "The user wants a detailed, well-structured essay of",
   "max_consec_line_repeat": 1,
   "degenerate_4gram_ge12": false,
   "tail": "nce length, output length, and the specific attention implementation used. Standardized benchmarks like MLPerf provide reference implementations, but long-context evaluation often requires custom setups.\n\nKey metrics include: prefill throughput (tokens/second for processing the input), decode throughput (tokens/second for generation, measured at various context lengths), time-to-first-token (TTFT,"
  }
 ],
 "verdict": "PASS"
}